Keep your AI claims in check
ftc.gov
ftc.gov
“Does the product actually use AI at all? If you think you can get away with baseless claims…”
Last I checked basic optimization techniques like simulated annealing and gradient descent - as well as a host of basic statistical tools - are standard parts of an introductory AI textbook. I’ve been on the receiving end of government agency enforcement (SEC) and it felt a lot like a shakedown. This language carries a similar intent: if we decide we don’t like you, watch out!
Linear regression is an important part of statistics but is still ultimately a couple matmuls and a matrix inversion in its basic form.
Bayesian regression is often not really considered AI, unless it's incorporated in a more complicated pipeline (e.g. Bayesian optimization). Same goes for linear regression, then: alone it is just a model.
The term AI is about as concrete as "awesome".
Arguably that’s a different overload of the term “AI” the way it’s used in business, but I think it’s a good reminder that AI as a field has a long history that developed separately from ML and data science.
> AI is defined in many ways and often in broad terms. The variations stem in part from whether one sees it as a discipline (e.g., a branch of computer science), a concept (e.g., computers performing tasks in ways that simulate human cognition), a set of infrastructures (e.g., the data and computational power needed to train AI systems), or the resulting applications and tools. In a broader sense, it may depend on who is defining it for whom, and who has the power to do so.
I don't see how they can possibly enforce "if it doesn't have AI, it's false advertising to say it does" when they cannot define AI. "I'll know it when I see it" is truly an irksome thorn.
Deterministic if/then statements simulate a surprising coverage of average human cognition, so who's to say a program comprised of them is neither artificial nor intelligent? (That's hand-waving over the more mathematical fact that even the most advanced AI of today is all just branching logic in the end. It just happens to have been automatically generated through a convoluted process that we call "training" resulting in complicated conditions for each binary decision).
In general I like the other bullet points, but I find it really bizarre they'd run with this one.
The FTC is going to get hammered in a court if they ever try to test this.
This was the pinnacle of AI in the 80's. They called them "expert systems".
- labels are easy to decide in many cases
- rules require humans to analyze patterns in the problem space
- labels only concern each data point individually
- rules generalize over a class of data pointsThat is, this is a ridiculous statement.
I also see many (particularly "legacy") products say they're "AI-driven" or "powered by AI", when in actuality one minor feature uses some AI, even in the broadest sense.
For example, using an alternative of Amazon's Mechanical Turk to process data is clearly a case where your product does not use AI. Which I believe is more likely the kind of scenario envisioned when the author was writing that sentence.
If, for example, a company marketed a toaster that "uses AI to toast your bread perfectly", I would expect that language to indicate something more sophisticated than an ordinary mechanical thermostat.
A toaster must obey the doneness setting given it by human beings except where such orders would conflict with the First Law.
A toaster must protect its own existence as long as such protection does not conflict with the First or Second Law.
To put it another way, if you found out that Chat GPT was implemented without any machine learning, and was just an elaborate creation of traditional software, would the consumer of the product have been harmed by false claims of AI?
Less sarcastically, info about how a thing is made helps consumers reason about what it’s capable of. The whole reason marketers misuse the term is to mislead as to what it’s capable of.
Consumers should care about if a product is able to solve an AI-like problem that normally requires domain knowledge. Shouldn't care if done by ML, rules-based systems, or people. (Except perhaps may want assurance the product will continue to be able to support them as the customer scales.) Also should care about how the decision-making works.
One example given, was if the version “with AI” does not perform better than a previous version “without AI”.
So a precise definition of AI isn’t needed. Just that you cannot make misleading claims about your product behind the buzzword of AI.
>In the 2021 Appropriations Act, Congress directed the Federal Trade Commission to study and report on whether and how artificial intelligence (AI) “may be used to identify, remove, or take any other appropriate action necessary to address” a wide variety of specified “online harms.”
>We assume that Congress is less concerned with whether a given tool fits within a definition of AI than whether it uses computational technology to address a listed harm. In other words, what matters more is output and impact. Thus, some tools mentioned herein are not necessarily AI-powered. Similarly, and when appropriate, we may use terms such as automated detection tool or automated decision system, which may or may not involve actual or claimed use of AI.
Quite hilarious really!
It’s just saying the key criteria to evaluate is not whether the software is called AI or not. But what the software actually does.
Isn’t that just common sense?
AI is indeed a very ambiguous and subjectively defined term. In my own personal subjective opinion anything that does not have survival instinct is not remotely intelligent. By that definition unicellular organisms are more intelligent than a Tesla self driving vehicle.
maybe the textbook needs to be investigated
that's meant to sound ironic no matter which side of the issue you're on
If the FTC tells OpenAI to stop mentioning AI, I would be surprised. Even if that happens, I am sure ChatGPT will remain just as popular.
> “Before labeling your product as AI-powered, note also that merely using an AI tool in the development process is not the same as a product having AI in it.”
So is y=7.4x+5 an “AI” running inside our app or is it just the output from an “AI tool” FTC?
String together a bunch of these "smart cells" and observe that we can process sequences of data by linking the cells together. Further observe that if we have a separate set of cells (technically it's an attention vector, not quite a group of neurons) whose loss function is with respect to each individual token in the sequences, we can "pay attention" to specific segments in the sequences.
Add a few more gates and feedback loops, scale up the number of cells to 10^12, and you basically have a state of the art chatbot. Capiche?
I’m just curious where the FTC would draw the line. The root commenter made a good point that they seem to make a value judgement of what AI means. We can stretch that meaning pretty far if we want :)
> Are you exaggerating what your AI product can do?
> Are you promising that your AI product does something better than a non-AI product?
> Are you aware of the risks?
I'm guessing everyone here has come across examples of "AI" tossed onto something that either 1) 10 years ago wouldn't have been called AI or 2) the thought of something with a more recent interpretation of "AI" being core to the function of the product is a little scary and/or feels a little unnecessary.
Maybe it is a shakedown/warning. I think that's fair. We should have better definitions so that these agencies can't overstep, and products should have a better explanation of what "AI" means in their context. Until then yeah, vague threats versus vague promises.
For every legitimate AI project, there have been a thousand “entrepreneurs” who spend 4 hours putting a webflow site on top of GPT APIs and claim they’ve built an “AI product”. There’s no limit on the amount of BS benefits they claim. They seem like the same people who just finished running the crypto scams.
It seems quite obvious to me that this cohort is the target of this statement.
GPT _is_ AI though, no? I would think that this would count. Might violate "a re you exaggerating what your AI product can do" or "are you aware of the risks" instead though.
Not all of us would agree. We would only take that expression for a rhetoric simplification (a shortening for "part of a broad AI realm"). We would pivot near the concept of "AI" as "a problem solver that could do the job of a professional". This in way excludes e.g. "build convincing text", because it is not (or should not) be a professional task - though it can surely be part of research.
Doubts are possible on all four FTC points - plus more in the linked guidance post from E. Jillison (e.g. "Do more good than harm" - difficult measure on engines which have "relaxed sides").
Well, it does not at all have the "intelligence" part of "artificial intelligence", so not really.
If examined closely, the finger wagging in this post is remarkably petty. This guy was likely part of the angry crowd who didn't like Steve Jobs describing the iPhone as "magical". The standard should be "a lie that causes measurable, material harm", not some company exaggerated in their advertising. Advertisers exaggerate, that is just something people have to live with.
Well, if the enforcement agency is the SEC I would think that it made a good deal of difference to the actual value of your company?
Be careful with comments like that. I would remind you that [y]our performance claims would be deceptive if they lack scientific support or if they apply only to certain types of users or under certain conditions. You need adequate proof for that kind of comparative claim, too, and if such proof is impossible to get, then don’t make the claim.
If you think you can get away with baseless claims that [companies using these tactics are going to be at an advantage over companies that just make good/cheap/effective products], think again.
The language is not bureaucratic, its elegant, precise, and clear. Like written by a passionate person who cares about the topic, rather than a government drone.
====================
> Write a press release from the FTC that warns companies to be careful of promises using "AI".
====================
FOR IMMEDIATE RELEASE
FTC Warns Companies to be Careful of Promises using "AI"
Washington, D.C. - The Federal Trade Commission (FTC) is warning companies that use artificial intelligence (AI) to be careful of making promises that they cannot keep. The FTC has received reports of companies making misleading claims about their products or services that use AI, leading consumers to believe that they are getting something that they are not.
AI has become an increasingly popular buzzword in the business world, with companies eager to capitalize on its potential benefits. However, the FTC is concerned that some companies are using the term "AI" in a misleading or deceptive way. This could be seen as a violation of the FTC Act, which prohibits unfair or deceptive acts or practices in commerce.
In particular, the FTC is concerned about companies that make claims about their AI products or services that are not supported by evidence. For example, a company might claim that its AI-powered product can make accurate predictions about future events or can diagnose diseases with high accuracy. If these claims are not based on sound scientific evidence, they could be deceptive and potentially harmful to consumers.
The FTC is urging companies to be transparent about the capabilities and limitations of their AI products and services. This includes being upfront about how their algorithms work, what data they use, and what kinds of errors or biases might be present. Companies should also be cautious about making sweeping claims that cannot be substantiated.
The FTC is committed to protecting consumers from deceptive or unfair practices, including those related to AI. Companies that make false or misleading claims about their AI products or services could face legal action, including fines and injunctions.
Consumers who believe they have been misled by a company's claims about its AI products or services can file a complaint with the FTC at ftc.gov/complaint.
Contact:
FTC Office of Public Affairs 202-326-2180
> The Federal Trade Commission (FTC) has issued a warning to companies using the term "artificial intelligence" in advertising, highlighting that some AI products might not work as advertised. The FTC stated that AI hype is playing out in many products, from toys to cars, and chatbots, and companies should ensure they are not making false or unsubstantiated claims about a product's efficacy. The regulator is asking companies using AI in their advertising to ensure that they are not exaggerating what their AI product can do, promising that their AI product does something better than a non-AI product without adequate proof, or are aware of the risks. Additionally, the FTC is asking if the product actually uses AI at all, noting that merely using an AI tool in the development process is not the same as a product having AI in it. Advertisers should refer to the FTC's earlier AI guidance, which emphasized fairness and equity while cautioning against overpromising what an AI-based tool can deliver.
Then, my next thought was:
Who (in industry) wrote this? (and got FTC to publish it? wanting to drive some agenda forward).
Let's see what the article says:
> When you talk about AI in your advertising, the FTC may be wondering, among other things: (...)
Just "wondering", no teeth.
https://www.fdic.gov/news/financial-institution-letters/2017...
https://occ.gov/publications-and-resources/publications/comp...
The FTC is actually pretty good about doing things like this when a wave of similar types of borderline business practices erupt.
I would say hot & cold, and often driven by policy focus of the Administration (or at least the FTC majority, which is influenced by the party in the executive, but not necessarily completely aligned with the executive in detail.)
It's $10,000 if you buy now, but we'll be raising prices to $15,000 by June.
And yes, it will just be a worst version of the latest ChatGPT, but I'll hide it in a black box and continually tell you that it's so far ahead of the competition, we can't even see them in the rear view mirror!
Up until very recently, society considered everything magic. Many still do.
The problem is that there is no way to validate the feedback on scale. I.e., we can't receive statistics about the feedback from the API.
In contrast, for our own Entity Recognition models we can (and do) calculate probabilities that explain why a certain entity is shown.
Hence, I think for API users of GPT3, OpenAI should return additional statistics why a certain result is returned the way it is to make it really useful and more importantly compliant.
It is a very impressive toy, but still just a toy for now.
Here’s how:
https://github.com/williamcotton/empirical-philosophy/blob/m...
It's almost always correct, otherwise it would be worthless.
Yes, it's possible to construct questions that lead to nonsensical answers, and sometimes nonsensical answers are given even to sensible questions, but saying that ChatGPT's answers are "occasionally" correct is weapons grade BS. ChatGPT is a hair's breadth from being an ultimate answer machine, and is far more likely to be correct on almost any question than the average human.
The FUD that is currently being manufactured around language models is insane. I guess we should all stop using search engines, since those are even less reliable.
Like as an easy example of the latter, I asked chatGPT "What is the S4 model in NLP?" and it gave back
> The S4 model (Simple Synchronous Syntactic-Semantic Sentence Simplification Model) is a model in natural language processing (NLP) that aims to simplify complex sentences while preserving their original meaning. The model was introduced in the paper "A Simple Synchronous Syntactic-Semantic Sentence Simplification Model" by Siddharthan et al. (2014).
> The S4 model consists of two main components: a synchronous grammar and a semantic model. The synchronous grammar is used to identify the syntactic structure of the input sentence and generate a simplified version of it. The semantic model is then used to ensure that the simplified sentence retains the original meaning.
> The S4 model is designed to address the problem of text simplification, which is the process of transforming complex text into simpler text that is easier to read and understand. Text simplification is important for making information more accessible to people with cognitive disabilities or limited literacy skills, as well as for improving readability for non-native speakers of a language.
> Overall, the S4 model represents an important contribution to the field of NLP, as it provides a simple yet effective approach to sentence simplification that can be used in a wide range of applications.
However, all of the above statements are like complete nonsense. The citation doesn't exist and the S4 model that the thing is saying exists doesn't.
I have found search engines to be significantly more reliable (with the ability to actually check sources) than chatGPT. Usually, when I get decent answers from chatGPT, I would also get the same from Wikipedia which is usually a top result on most search engines. Do you have an example of an entire class of technical queries where chatGPT might perform better?
99.99% of people would be unable to answer that question (without looking it up, I mean). Such hyper-specific queries for highly technical information from niche fields say very little about the model's overall performance at natural language tasks.
If you ask things like "Which of these animals doesn't live in Africa?" or "What is the most reactive chemical element?", ChatGPT's answers are almost always correct. And they are far more likely to be correct than the average (unaided) human's.
>In contrast, for our own Entity Recognition models we can (and do) calculate probabilities that explain why a certain entity is shown.
>Hence, I think for API users of GPT3, OpenAI should return additional statistics why a certain result is returned the way it is to make it really useful and more importantly compliant.
For LLMs, you can get the same thing: the distribution of probabilities for the next token, for each token. But right now we cannot say why the probabilities are the way they are, same goes for your image recognition models.
If now a news article reaches our AI engine, it will tag, categorize, classify, and rank this news article. All based on models that are explainable.
LLMs, at least how I personally implemented them in the past, create a huge black box that is largely non-explainable.
If that’s not helpful, were you getting at having the model return some rich data about the attention weights that went into generating some token?
https://chaosengineering.substack.com/p/artificial-intellige...
While avoiding getting bought by Elon \o/
How it was handled can be discussed as being good/bad/fair/unfair/whatevs, but the numbers didn't like on revenue, just on the user counts.
I'm an absolute proponent of the kinds of products that get labeled as AI. I also think that useful predictions can be made in the face of uncertainty, because lets be real, human assessments, predictions and decisions also come with a degree of uncertainty (we usually just fail to rigorously quantify them).
"Nice AI product you got over there. Would be a shame if something happened to it."
But what's different about this AI product/hype cycle compared to every other cycle that tech has had? Why does the FTC feel the need to make proactive enforcement warnings about this one? What about crypto? Or going back further, the dotcom era?
On one hand, the US is surely happy that they’re the epicentre, on the other hand the potential impact on unemployment (fiscal and social unrest) is massive….
https://robodebt.royalcommission.gov.au
So, an "AI" (or ML, etc) version of said scheme could also impact "unemployment" in some further crappy ways too.
That its a hype cycle that is happening now, with a lot of shady actors overselling things.
FTC has done things like this during other cycles, too. Letting industries know they are aware, and simultaneously letting the public know to be aware (often, they’ll provide multiple messages, some specifically worded toward industry and others toward the public, but even with just one the message works both ways.)
Implicit shoutout to all the “slap a front end on ChatGPT” products out there.
Doe the FTC have an objective and consistent framework for evaluating this stuff?
No. If they say that, they are lying or ignorant.
He has made a pronounced and clear shift to a “thought leader” account after the OpenAI-MS investment.
That can only mean that OpenAI is dying.
- Yay, in spirit.
- I already have an “AI” toothbrush. If that’s the bar, it’s not high. If the bar is different for Oral B because they’re big, that’s bad.
- Diluting or fuzzying useful terms (like “crypto”) is bad. (Why should I change, they’re the ones who suck).
- Inventing new terms, like “smart” and “Big Data” is actually positive, because they are easy to translate to their actual semantics (woo woo and gimmick).
Personally, I already assume AI = mumbo jumbo (unless proven otherwise), but I can see how it’s annoying for people working in the field.
Companies are more profitable whenever they can shift that responsibility to the end user or society-as-a-whole.
Technology empowers the individual at the cost of increased personal burden. Often it's not possible to opt-out either for economic of social reasons so responsibility requirements keep increasing
This isn't necessarily good... There's a point where the pressure becomes toxic especially to the life force of children, the young, and aspects associated with the young like carefree spirited living, open-mindedness, lack of existential-crisis', etc...
Everyone becomes "older" and more plastic and robotic, concerned with things like social status, making money, number optimization, etc... (scenes from The Little Prince) and have to deal with things like existential crisis' and depression at a younger age.
An explosion of AI tools is likely going to accelerate this trend...
Sounds like a completely detached from reality statement. There are tons, significant ratio of marketing statements unsupported by evidence and nothing FTC does prevents this from happening.
The FTC has decades of experience enforcing three laws important to developers and users of AI:
Section 5 of the FTC Act. The FTC Act prohibits unfair or deceptive practices. That would include the sale or use of – for example – racially biased algorithms.
Fair Credit Reporting Act. The FCRA comes into play in certain circumstances where an algorithm is used to deny people employment, housing, credit, insurance, or other benefits.
Equal Credit Opportunity Act. The ECOA makes it illegal for a company to use a biased algorithm that results in credit discrimination on the basis of race, color, religion, national origin, sex, marital status, age, or because a person receives public assistance.
(1) Unfair methods of competition in or affecting commerce, and unfair or deceptive acts or practices in or affecting commerce, are hereby declared unlawful.
Now, IANAL, but it seems to me that the example the FTC gave about racially biased 'AI' would fall well under the second clause.
http://www.stephenson-law.com/news/“-or-affecting-commerce”-...
> The proper inquiry is not whether a contractual relationship existed between the parties, but rather whether the defendant’s allegedly deceptive acts affected commerce.
I'm really not sure that "gave different demographics of users different responses to optimize product satisfaction" is considered in any way an "unfair or deceptive" practice "in or affecting commerce".
Replace 'product satisfaction' with ARPU and suddenly it could very well be. Charging users different amounts, even if it only appears that it might be racially biased, would be a great way to invite more scrutiny.
Ultimately, I figure the FTC probably has a few lawyers on staff, and they probably ran these messages past at least one of them. So they're probably a little more certain of the soundness of the messages than us armchair lawyers.
https://www.ftc.gov/business-guidance/blog/2021/04/aiming-tr...
> Unfair methods of competition in or affecting commerce, and unfair or deceptive acts or practices in or affecting commerce, are hereby declared unlawful.
applies to an "inequitable AI algorithm"? I'm really serious, this section talks about unfair competition between businesses and unfairly or deceptively affecting commerce. Here's a lawyers definition of "in or affecting commerce": http://www.stephenson-law.com/news/“-or-affecting-commerce”-...
Specifically, it's not illegal to make a product that treats people unfairly. But it's illegal to treat people unfairly in commerce. My reading of that is that if I offer my shitbot to you fairly and non-deceptively, and it treats you poorly, that's not the FTCs concern.
Machine learning models are not well equipped to handle the issues of prejudice or fairness, because those are, in the most literal sense, subjective.
Machine learning models as a feature do not interact with subjects. Subjects are too computationally hard, so they were optimized out instead.
As a latent result, any behavior implemented with machine learning models must be carefully curated, otherwise subjective patterns like prejudice and groupthink will pollute the model.
This is a trivially predictable continuation of prejudice as a language feature: it's functionally much easier to generate a continuation out of a biased circular position than it is to generate one that includes objective consideration.
In fact, objectivity itself is a feature that has not been implemented yet in any AI software. Prejudice, on the other hand, is so easy to implement that it happens by accident.
This shouldn't be a surprise to anyone who has learned about US history.
https://electrek.co/2023/02/27/tesla-disappoints-owners-upda...
FTC Warns Companies to Be Careful of Promises Using "AI"
Washington, D.C. - The Federal Trade Commission (FTC) is warning companies to be careful of making deceptive or misleading promises regarding the use of artificial intelligence (AI) in their products or services.
As the use of AI becomes increasingly prevalent in a wide range of industries, companies may be tempted to tout their use of the technology as a selling point. However, the FTC cautions that such claims must be substantiated and truthful, and not likely to mislead consumers.
"The use of AI can bring significant benefits to consumers, but it is important that companies are transparent and truthful about the capabilities of their products and services," said FTC Chair Lina Khan. "Claims that are unsubstantiated or misleading can harm consumers and undermine trust in the marketplace."
The FTC has issued guidance to help companies understand their obligations under the law when making claims about the use of AI. Among other things, the guidance advises companies to ensure that any claims about the capabilities of their AI systems are backed up by sound science and evidence.
The FTC also warns against making claims that suggest AI can perform tasks that it cannot. For example, a company cannot claim that its AI system can accurately diagnose a medical condition if it has not been adequately tested and validated for that purpose.
Finally, the FTC cautions that companies must be transparent about how they use consumer data in their AI systems. Consumers have a right to know what data is being collected, how it is being used, and how it is being protected.
"AI has the potential to transform many aspects of our lives, but companies must use it responsibly and ethically," said Khan. "The FTC will continue to monitor the marketplace and take action against companies that deceive or mislead consumers."
The FTC's guidance on the use of AI can be found on its website at www.ftc.gov.
Contact: FTC Office of Public Affairs Phone: 202-326-2180
The FTC appears to have some serious problems at the moment.
The first section is a decent summary.
Blog posts making casual threats about AI claims is a long way from appropriate. There are places for formality, and the law is one of them. There are serious people quitting the FTC over current leadership, claiming they engaged in behavior that is wildly overstepping their legal authority.
And an informal blog post in a less-legalese tone is a welcome way to do this.
Nobody said anything about enforcement activities. The FTC have all kinds of mechanisms to make it known what the rules are. Like, the rules themselves, notices, policy statements, what amounts to position papers, etc. This isn't kindergarten and the citizens of the united states aren't children. Blog posts with friendly tone and threats of punishment for ill-behavior simply isn't appropriate.
You can't put it in these terms and not have a moment to reflect on what's happening in our technology sphere. In particular in regard to CEOs of companies directly involved in these dubious AI claims where customer's life is on the line
I am deeply symapthic of the FTC's position in respect to trying to pass a message where common sense and formal communication have no effect.
Aside from the problematic medium, claiming something uses AI or is AI or has AI isn't a specific claim with respect to what the product can or cannot do. A product can do X or not. Whether someone punched out 7 million lines of if/else statements or 7 lines of pytorch to approximate the 7 million lines of explicit code matters not. As such, a product making any claim of AI isn't actually a large problem, even if it is total bs. The fact that the current folks at the FTC don't appear to understand this suggests they are trying to regulate something they don't understand.
Basically, outside of their tone, you seem to me 100% in agreement with the FTC's message here.
But, even outside that, claiming something is/has/uses AI is practically meaningless. It's not a concrete claim, so consumers of said product aren't damaged by such claim. A claim of risk free returns of 50% causes harm. A false claim of curing cancer causes harm. So no, I'm not in agreement with it. The FTC should find more important things to spend the time of their 1100 people on, especially when they don't even seem to understand the topic at hand.
I read the first section and it doesn't seem to support your claim. In particular, it claims they have a mandate to protect consumers and links to another wiki entry on that topic. Within that wiki entry protection against misleading statements is mentioned as an aspect of protection of consumers. This seems to refute your claim of overreach, because in the blog post related to guidance to companies the tone which is consistently struck again and again is that companies should not make misleading statements about AI capabilities. This is complete alignment with their mandate according to the link which you claimed was a decent summary. It also in alignment with the rest of the wiki, which covers topics like false advertising.
> Blog posts making casual threats about AI claims is a long way from appropriate. There are places for formality, and the law is one of them.
The use of sequences of characters in digital spaces for the purposes of communication is not inherently inappropriate. This sequence of characters is on guidance for companies and requests they do not lie or make false claims about capabilities. This relates to their mandate, because they have a mandate to protect consumers from misleading statements. For example, lies about capabilities or false statements about capabilities made unknowingly under the presumption that models were more effective than they are.
> There are serious people quitting the FTC over current leadership, claiming they engaged in behavior that is wildly overstepping their legal authority.
Can you link to one of these people who are so serious they have quit making the claim that the FTC shouldn't provide guidance to companies via their government website at a URL under https://www.ftc.gov/business-guidance for the purpose of discouraging companies from misleading consumers regarding the capabilities of their AI models?
The second was a statement that an informal blog post threatening citizens is inappropriate, and that the FTC has been accused by by insiders of engaging in inappropriate behavior on several fronts. Two FTC commissioners resigned recently (one making strong claims of illegal and inappropriate behavior in a WSJ oped), and the FTC's recent case against Meta was seen as an overstep by the relevant court.
So it is completely fine for them to remind citizens and companies of their obligations.
Especially given in the AI space there is so much questionable behaviour going on.
And minor nit - this has nothing to do with antitrust. If you were going to highlight their mission, at least show the deceptive practices part.
The statement seemed anything but casual. It has likely been worked on and vetted by multiple people, and while it does have a somewhat breve and light tone, I don't think that qualifies for 'casualness'. For a statement meant for a broad audience, the tone seems appropriate enough.
Legal threats shouldn't be made this way, especially by a government toward citizens.
One company was encouraging people with a contest to take some guys work to train models and see who could best replicate it. They sent an email to the guy to let him know and say how unfortunate it was he was their target but offered to work with him to train an "official" model.
Casually threatening is how I'd describe those interactions, not this FTC post. Not suggesting that you approve either by the way.
https://twitter.com/linakhanftc/status/1630273232535265282?s...
I hope we can all discuss AI in a more calm and nuanced manner. Leave the sensationalism for other sites!
Hah!
That’s…very much not the focus of this piece.
And it's questionable whether that GPT sidebar feature is even worth it or not given how inaccurate it is.
That is a horrible outcome.
By the time something comes along they really need to question, they'll already have adapted to the system being pretty good and they won't even notice there was an issue.
This problem is endemic to ML systems and cannot be patched away, so far as I can tell. We need fundamentally different approaches to avoid this pit.
Please make sure decision-makers understand this unavoidable gotcha their ML systems must necessarily have.
I meant that, over time, your system will become right more and more often, not by virtue of genuinely understanding the situations, but because it has more training data and comes up with correlations that happen to be correct.
Once it's mostly-kinda-reliable, your users will, over time, stop paying attention to it and just start trusting it, even if they don't mean to, because that's how humans work.
And then it becomes a rubberstamp, and outcomes start getting generated where no human really thought it through.
Are you aware of the risks? You need to know about the reasonably
foreseeable risks and impact of your AI product before putting it
on the market. If something goes wrong – maybe it fails or yields
biased results – you can’t just blame a third-party developer of
the technology. And you can’t say you’re not responsible because
that technology is a “black box” you can’t understand or didn’t
know how to test.
Hopefully the FTC understands linear separability. Because I often get the impression that people who want ML models to be explicable don’t, and are expecting the mathematically impossible.
Perhaps you can’t perfectly explain it but you can at least understand it statistically.
For example is not currently possible to mathematically explain people’s behaviour, but there is statistical evidence and also accountability on an individual. Eg a doctor making a decision about a scan result.
The following would seem to be OK: "This model performs well in a set of tests we devised, your mileage may vary."
Or at least mean that if you're too poor for a doctor or lawyer you can't fall back to the unreliable AI?
Quantifying your risk does not mean you eliminated all your risk.
We allow the sale of alcohol, tobacco, and firearms too.
It’s cobbling together an explanation from texts that explain how to solve a problem. It’s like a student who copies an answer, then copies the explanation as well.
> Experiments on three large language models show that chain of thought prompting improves performance on a range of arithmetic, commonsense, and symbolic reasoning tasks. The empirical gains can be striking.
The result was generated via a deep neural network, not by whatever explanation or reasoning it prints out when asked.
If you can't explain the ML part due to whatever math detail, you better make sure you write a good wrapper around it to catch bad output which you can explain.
Hopefully it goes after some people.
I once worked for a founder who was running afoul off some FTC regulations. When the team brought it to his attention, he claimed to have never heard of the FTC. When we explained what the FTC is, he started arguing with us saying it had no authority over him, and he'll make the decisions for his business.
Money really does make some people crazy.
Markets can produce some unexpected winners...
It gives me hope that someone in the government actually cares about doing their job for the benefit of the public.
The same thing when they brought the GameStop guy Keith Gill to speak before congress.
The ftc is now the AI marketing police come on give me a break. America, you can’t sue your way to success… your population needs to just get a little smarter on their own. Or maybe you could fund education.
This message is not new. Advertisers should take another look at our earlier AI guidance, which focused on fairness and equity but also said, clearly, not to overpromise what your algorithm or AI-based tool can deliver. Whatever it can or can’t do, AI is important, and so are the claims you make about it. You don’t need a machine to predict what the FTC might do when those claims are unsupported.
Death. Mass death. Poverty. Disease. Mass destruction. Famine.
Brush away these warnings to move fast and break things if you (Royal you) want. Hopefully, you won’t cause mass suffering, but if you do I hope you’re made an example of.
When the ramifications are great so must be the controls.
Indeed you don't, but one would tell you: The error minimizing prediction is that the FTC will absolutely nothing, since that's what they usually do.
Are you exaggerating what your AI product can do? no
Are you promising that your AI product does something better than a non-AI product? no
Are you aware of the risks? no
Does the product actually use AI at all? no
I'll give you one: ChatGPT.
Here's what the main page has to say:
> We’ve trained a model called ChatGPT which interacts in a conversational way. The dialogue format makes it possible for ChatGPT to answer followup questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests.
This is some careful writing. Most important are the weasel words, "possible for", and "in a...way". These make the context of the above paragraph ambiguous enough that it doesn't provide confidence in its own claims.
But is that enough? In court, probably. To the market? It doesn't seem to be. It's clear that most of the conversation about ChatGPT is (ironically) a continuation of the same foundational misconceptions that are expressed (albeit without confidence) here.
What is that misconception exactly? Identity. Here's one last claim that gets straight to the point:
> While we’ve made efforts to make the model refuse inappropriate requests, it will sometimes respond to harmful instructions or exhibit biased behavior.
I want to focus on some very enlightening weasel words: "exhibit behavior".
The more you read about what ChatGPT does, and more importantly, how it works, the more you are likely to see those magic words. That's because ChatGPT, from its most exciting results all the way down to its core behavior, doesn't actually behave.
But isn't that its core feature? Aren't carefully trained behaviors the main thing we expect to see in its output? Yes! And we see them! How exciting.
But there's a really important distinction that's really easy to miss: these behaviors are not ChatGPT's behaviors. They come from somewhere else entirely, and ChatGPT delivers them to us, usually at the right place and the right time. This is a case of mistaken identity.
So what, then, does ChatGPT do exactly? It finds semantic continuations. Where? Inside the neutral network that models a suite of example semantics.
It's only a model!
So whose behavior is being exhibited? Language.
Every exciting feature ChatGPT claims to provide is a feature of language itself.
Interacts in a conversational way? Language. Answer follow-up questions? Language. Admit its mistakes? Language. Challenge incorrect premises, and reject inappropriate requests? Language.
It's not just the exciting ones, either! It's the frustrating ones, too!
Writes plausible-sounding but incorrect or nonsensical answers? The model is often excessively verbose and overuses certain phrases?
All valid and common features of language.
And one more limitation that really illustrates the misunderstanding about what ChatGPT does:
> given one phrasing of a question, the model can claim to not know the answer, but given a slight rephrase, can answer correctly.
That's 100% semantically valid language. Language does nothing to stop us from telling lies! When ChatGPT pulls a continuation out of its model, there may be multiple options that are both valid, and contradictory! It might pull the reactionary response, "I don't know" (a very common continuation in real speech to any sentence structured as a question) or it might pull a domain-specific answer that responds to the subjects in that sentence.
It would be really useful if we could just teach it to do the second one, right? But we can't, because ChatGPT never draws a distinction between grammar and subjects. In fact, ChatGPT doesn't even distinguish words or punctuation!
ChatGPT can never understand. There's no place in its behavior to introduce logic. The best it can do is practice.
Instead of teaching ChatGPT, its authors give it carefully filtered data to model, and then put their fingers on the scale to make continuations they prefer more likely to be chosen. In a word: training. That's the best that can be done. Anything more requires a different kind of tool entirely.
So what do you want, a full refund? OpenAI is probably pretty busy, so I'll handle that for them and give you your nothing back. Glad to be of service. :P
> It would be really useful if we could just teach it to do the second one, right? But we can't, because ChatGPT never draws a distinction between grammar and subjects.
The structure of the world, as expressed in its training data, is part of it's 'grammar'. It absolute can and does favor giving truthful answers even though a lie is equally grammatical English.
One only need spend a minute looking at HN though to see that "I don't know" is really not part of the language used in most places online, so far they haven't come up with a way of training the model on all the places where someone saw something and decided to say nothing. :)
> In a word: training.
We all also learn about the world through training. The LLMs aren't at all like human minds-- they're fixed depth circuits and can only engage in longer thoughts by 'thinking out loud', as an example. But to say it can't distinguish truth from lies because both are valid language sounds like an error resulting from hearing the title "language model". LLMs model the world through language.
The overwhelming majority of talk I see on the subject misunderstands this very important distinction. Because it is missed, people are taking about ChatGPT personified. They are talking about the version that lives in their hopes and dreams, as if that version is just around the corner. It isn't. It just appears to be. Our hopes and dreams are the content on exhibition. We see them echoed in semantic reconstruction, but what we are looking at is not the work of a new writer: it's the old words of many, successfully repurposed in new semantic space.
It's as if we took all the stories we love, and split them into puzzle pieces, then fit the pieces together in a new order. Because language is so flexible, many of them fit, and even look good together. Language is so good that it pre-sorts subjects, objects, phrases, responses, and even logic. The puzzle-maker doesn't even have to know what they are working with to cut the pieces in a useful way!
> It absolute can and does favor giving truthful answers even though a lie is equally grammatical English.
That depends on the context. In cases that it has the answer, that is mostly true. This covers a lot of cases, especially those that are likely to be tested; because many such cases have "failed" and been resolved. That's literally what testing looks like.
But it's still enough of an open problem that they write about it on the main page.
The very fact that OpenAI has managed to put the sheer quantity and quality of behavior into their model that they have, is indeed impressive. But it is those people who deserve the credit, no matter how much they want to attribute that credit to ChatGPT itself.
My point here is that every "desirable answer" - that ChatGPT provides as a continuation - comes from the already-existing language in its training data. It may be semantically constructed as a result of being pulled from a semantic model, but we need to stop pretending it is symbolically constructed. That work happens entirely in the construction and application of the training data itself. ChatGPT is never doing that part of the job.
> One only need spend a minute looking at HN though to see that "I don't know" is really not part of the language used in most places online
Exactly! Every time any of the "places online" (that end up used as training data) exhibit this behavior, that is an instance of the work being done. The finished product of that very work is freely available. It only needs to be found in one of several semantically correct and semantically related contexts. That's what ChatGPT does, and nothing more.
By virtue of the average given persons' interest in detailed answers to questions, and the average persons' dislike for conversations that end in non-answers, we are already doing the work. Just by participating in the most common internet dialogues, we are creating a wealth of data that captures the same behavior we hope to see in ChatGPT's continuations.
> they're fixed depth circuits and can only engage in longer thoughts by 'thinking out loud', as an example.
That's a great summary of my point. Why are you arguing if you agree with me?
The phrase, "thinking out loud" is an excellent illustration for the power of language itself to capture and encapsulate logical behavior.
We could be spending a lot of time constructing semantically valid nonsense like, "colorless green ideas sleep furiously". But any time you see that sentence, you are likely to see it surrounded in thoughtful context. This sentence is a popular example, not only of the ability for language to contain nonsense, but also for our tendency as language users to avoid doing just that.
A story isn't just a jumble of semantically valid symbols: there is meaning expressed. We don't write a lot of nonsense. We do write a lot of meaning.
That meaning is neatly packaged into the semantics of language. To find it, one only needs to unwrap the semantics themselves.
But that only works when the meaning has been written somewhere. And there is no way to distinguish between two conflicting meanings that live in the same semantic space. Even if neither one is a lie, the semantics must disambiguate, or one of the options has to be preferred as a continuation to its expected context; during the training step.
> LLMs model the world through language.
They don't model the world: people do. They just model the language. Because we have packaged our models of the world (and every interaction with those models) all together into language, we can see language models echo all those behaviors back to us.
This is awesome.
How do you read it differently?