A view from DC: Consumers deserve the truth about AI systems

Last week, the U.S. Federal Trade Commission issued a request for public comment on a proposed policy statement regarding artificial intelligence accuracy, advancing a novel theory that developers could violate deception prohibitions under Section 5 of the FTC Act if they tune models to achieve "undisclosed ideological objectives." Comments are due by the end of July.

Another recent development: the trailer for the final installment of the film adaptation of "Dune" came out yesterday. Please bear with me.

Truth is the mind-killer?

You may or may not recall, depending on how much of a sci-fi nerd you are, some of the backstory to Frank Herbert's vast universe of sand worms and mystical secret societies. The novels take place long after a revolt known as the "Butlerian Jihad," which resulted in the eradication of all thinking machines across the human-colonized universe. I won't go into why the eradication of all automated decision-making was seen as necessary, but suffice it to say a few catastrophic risks had been realized over the centuries.

No doubt slightly mourning the loss of their helpful AI companions, humanity decided to invent biological replacements that could still handle the computational density required to run an interstellar empire. Enter the Mentat discipline. In lieu of computers smarter than humans, a special class of superhumans emerged, conditioned to function as organic computational engines. If you've seen the recent "Dune" films, you may recall a character — portrayed by Stephen McKinley Henderson — whose eyes roll back in his head and go all creepy and white before answering a question. That's Thufiir Hawat and he's a Mentat for House Atreides. 

Anyway, Mentats act much more like a traditional machine learning model than a large language model. They are not stochastic parrots. Instead, they ingest vast troves of data and synthesize it through an ingrained logical framework to flawlessly calculate probabilistic outcomes, helping their masters make sense of reality, and thus make better predictions and decisions. 

Nevertheless, one lesson Herbert brings home repeatedly in his novels is the fact that a computational system is only as good as its inputs. Mentats repeatedly fail to deliver accurate counsel because, despite their best efforts, they frequently rely on incomplete, poisoned or culturally biased data.

The repeated failure of Mentats is part of a broader theme in Herbert's work. Not only is objective truth something that can never be fully attained, but the Dune universe also paints the pursuit of truth as fundamentally flawed. Whether through the depiction of the Bene Gesserit religious order engineering a local culture's understanding of reality or the realization that complete prescience of future events is an inescapable prison, Herbert pushes his readers to understand that truth is a tool of control.

In his telling, to believe in a single source of truth is to be stuck in a cognitive cage that, more often than not, is imposed by some outside force.

Back to the plot: Consumers expect AI accuracy

Under Chairman Andrew Ferguson, the FTC's open request for comment pursues the White House's December 2025 executive order, which explicitly directed the FTC to target state laws requiring the alteration of "truthful" AI outputs.

Reflecting on this executive order and its instructions to the FTC, I previously wrote about the multifaceted nature of "truth," both as a philosophical ideal and as a practical goal.

The proposed policy statement "concerning the suppression of accuracy in artificial intelligence systems" is short and relatively straightforward. The agency warns against the subversion of AI systems in any way that would create a deviation from underlying consumer expectations that generative AI outputs should match reality.

In so doing, the FTC makes the assertion that generative interfaces, as they have been presented to consumers, have made an explicit or at least implicit claim toward truth and accuracy in outputs. 

One might wonder where the yardstick should come from. What is the measure of truth and accuracy against which we should govern our AI systems? In policy statement framework, "truth and accuracy" might mean a model faithfully reflects its underlying training data, even if that data produces skewed or biased results. Or it could mean training data sets should be as reflective of reality as possible at the first instance, to help ensure accuracy in outputs.

The draft policy statement also recognizes that accuracy is not the end-all of every chatbot interaction. "Users of an AI chatbot might, for example, reasonably expect the system to balance succinctness, clarity, relevance, accuracy and other objectives in its attempt to produce the best output. And while truth and accuracy are in many cases implicit objectives in requests to AI systems, a user could request output that is intentionally inaccurate or that deprioritizes accuracy in favor of some other objective like entertainment.”

Ultimately, however, the FTC hopes to put AI developers on notice that any "unexpected objectives" or "hidden agenda" underlying their models performance must be disclosed clearly and completely to properly set consumer expectations, even if required by another law.

Colorado — still — in the crosshairs

As ordered to do, the FTC also takes aim at state laws, arguing that laws imposing liability for discriminatory outcomes incentivize AI companies to falsify or artificially steer their outputs to balance the scales, thereby deceiving consumers. As the White House previously did, the FTC explicitly names the Colorado AI Act, now repealed and replaced. The proposed policy statement claims the repealed law would have created "a broad duty on AI companies to avoid output that might lead to disparate impacts in various contexts, including when a customer’s foreseeable use of that output could itself create a disparate impact." 

The "broad duty" in question only would have applied in the limited circumstance where an AI system was a substantial factor in making a "consequential decision" regarding employment, housing, financial lending, healthcare, education or essential government services. 

The now diminished replacement law in Colorado receives similar treatment under the proposed statement because it extends liability for discriminatory outcomes to developers. The FTC states, "It is predictable that an AI company might suppress accuracy and interpose other objectives, such as so-called 'equity,' to avoid liability under this law, but fail to disclose these ulterior objectives in order to hide the loss of accuracy they necessitate."

Is it time for a truth campaign?

Across the political landscape, stakeholders agree there is a crisis in consumer expectation of how generative AI systems work. All seem to agree, too, that the public statements of AI developers have helped to set these expectations.

In a recent piece for Tech Policy Press, for example, two former attorney advisors to FTC commissioners, Gaurav Laroia and Charlotte Slaiman, argued there should be a massive public interest messaging campaign around AI to properly reset consumer's understanding of how generative systems work. Echoing the FTC, they cite the high reliance consumers place on these outputs and claim developers have no incentive to fix this, writing, "We are currently fighting an asymmetric battle on messaging with the companies."

Rather than focusing on policing the output of algorithms, Laroia and Slaiman argued the priority must be on cultivating digital literacy that demystifies the predictive nature of these models. Their solution would mirror the public health frameworks utilized in the Tobacco Master Settlement Agreement, with money flowing from the types of settlements we are seeing around product liability.

Ultimately, all seem to agree that consumers are finding themselves lost in AI hype and deserve clarity on how these systems operate. Machine omniscience is no less fallible than its human counterpart. The solution is probably a mix of multiple approaches, including robust model governance, accurate marketing claims and even public messaging campaigns. 

Interstellar conflict might not yet be on the table, but the stakes are still high.

Please send feedback, updates and spice to cobun@iapp.org.