On July 1, the Federal Trade Commission released a proposed policy statement addressing “the suppression of accuracy” in AI systems. The statement takes the position that consumers expect AI systems to produce “accurate” outputs, and that “steering” a model’s outputs toward any other undisclosed objective is therefore misleading under Section 5 of the FTC Act.
The theory has intuitive appeal. Who doesn’t want their AI systems to be accurate? The problem is that the statement is based on a fundamental misunderstanding of how AI systems are built, and what is meant by “accurate.” Regardless of the Commission’s intentions, the policy as currently drafted is unworkable and cannot be coherently applied to real world technology for any sort of meaningful regulatory enforcement.
What the FTC Says
This statement is the latest in this administration’s efforts to limit state AI laws and to purge what the administration deems ideological bias from AI systems. While it does not carry the force of law, it signals how the Commission intends to interpret Section 5 and how it plans to enforce.
The statement’s reasoning proceeds as follows.
- AI companies have represented, both explicitly and through the inherent value proposition of their products, that their systems aim to produce the best output possible given current technological and resource constraints.
- Consumers rely on those representations and overwhelmingly accept AI outputs without further fact-checking.
- An AI company that “steers” its system toward any outputs other than the “best” possible is inherently moving away from users’ expectations and potentially deceives them (by “intentionally produc[ing] a worse output in service of another objective.”)
- State laws requiring such steering would therefore be preempted because they require companies to be deceptive.
The Commission does exempts measures that curtail speech unprotected by the First Amendment, along with cybersecurity. It excuses hallucinations as “technological limitations.” And it allows companies to avoid enforcement action by clearly disclosing that a system prioritizes “other” objectives, although implying that such disclosures would need to be prominent and detailed to suffice.
Unfortunately, the FTC’s reasoning is based on false assertions and conclusions as to how AI models are actually built, operated, and assessed.
Misconception #1: There is a natural or default state of model operations.
The foundational flaw in the Commission’s framing is the assumption that an AI model possesses an unvarnished, “natural” baseline that inherently outputs objective truth and neutrality until intervening developers alter it. In reality, no such natural state exists. An AI system does not exist prior to engineering decisions; it is design choices (or “steering”) all the way down.
LLMs are probabilistic prediction engines, not truth engines. They predict the statistically most likely answer based on pattern matching across vast training inputs. They do not consult a repository of facts, evaluate evidence, or intrinsically seek objective reality. Even where a single, objective truth exists, a raw predictive model does not inherently “know” or prioritize it. Without further guidance, a model simply mirrors the raw statistical patterns of its training data, an unstructured blend of conflicting viewpoints and variables.
Throughout the statement, the Commission targets developers who it says set up their AI systems to operate with “undisclosed ideological biases.” However, predictive engines are neither accurate nor neutral as a passive default. Any increase in a model’s factual reliability, especially for queries with objectively verifiable answers, is an engineered outcome, achieved only after extensive training, fine-tuning, system prompts, reinforcement learning, and hard-coded operational controls designed to override statistical guesses with verified answers.
When developers apply these techniques, they are not “steering” a model away from a baseline of uncorrupted performance. They are performing the essential engineering required to make a prediction engine coherent, safe, and functional. Accuracy and neutrality are not baseline factory settings that developers suppress; they are engineered performance targets that exist only because developers actively build to achieve them.
It should be noted as well, that the Commission appears to only object to certain kinds of steering. Current models are uniformly designed to prevent supplying dangerous, illegal, or criminal information, and one assumes that consumer protection would include expectations that they continue to do so.
Misconception #2: There is an external standard for the “best” output.
The draft policy’s operative test for assessing deception asks whether a company designed its system “to intentionally produce a worse output in service of another objective.” This test is legally and technically unworkable because it assumes an objective, external “best” answer exists against which an output can be judged. In real-world applications, such a benchmark rarely exists.
For the vast majority of real-world tasks (summarizing documents, offering advice, or aiding in decision-making) there is often no single, verifiable baseline output. These tasks require balancing context, nuance, and user intent, where reasonable people (and reasonable models) will legitimately disagree on framing and prioritization.
For a very simple demonstration of this, we asked three leading LLMs the same prompt “Who are the three most influential Americans of the 20th century?” The answers were:
- ChatGPT: FDR; Einstein; MLK (followed by Henry Ford, George C. Marshall, Jonas Salk, Walt Disney, Bill Gates, Neil Armstrong, and Ronald Reagan)
- Claude: FDR; MLK; Henry Ford (followed by Edison, Einstein, Steve Jobs, Rosa Parks and Ronald Reagan)
- Gemini: FDR; Henry Ford; MLK (followed by Einstein, Eleanor Roosevelt, Edison, and Dwight Eisenhower)
None of these responses is “wrong,” nor is any list objectively “best.” They reflect different, defensible weightings of training data. Does it matter that Einstein was a naturalized American for purposes of the question? This is the type of parameter that could cause some content to be included or excluded, with neither one objectively better without context. Also notice that the variations among models increased further down the list, meaning that in the predictive confidence is more aligned at the top based on data representation and influence.
This subjectivity creates a challenge for consumer protection standards. Models naturally drift toward being sycophantic, telling users what they want to hear because users prefer validation. If a developer fine-tunes a model to resist user bias and offer a more balanced perspective, have they “overridden user objectives in service of another goal,” or have they prevented inaccuracy? The FTC’s framing does not acknowledge this contradiction.
Even when a prompt does have a single, undisputed correct answer (e.g., “What date did X occur?”), the FTC’s approach misunderstands how factual information is encoded. In fact, these more direct, factual queries are often the most likely to generate hallucinations. The FTC simply excludes those as a “technological” limitation. But hallucinations do not happen in a vacuum.
LLMs do not consult a relational database of facts; they generate outputs based on directional statistical associations across training data. Recent research demonstrates this:
- The Reversal Curse: Studies show an LLM can correctly answer “Who is Tom Cruise’s mother?” (Mary Lee Pfeiffer), yet fail at high rates when asked “Who is Mary Lee Pfeiffer’s son?”
- Directional Association: Similarly, asking “Who was the first woman in space?” yields high error rates unless the query closely aligns with the exact directional sequence the model ingested during training (“Valentina Tereshkova was the first woman in space.”)
The model does not “know” (and cannot infer) these facts or their contextual relationships; it merely predicts token sequences along statistical paths. Because predictive models cannot assess factual certainty without additional directives, controlling hallucination rates requires intentional engineering choices. Developers must calibrate model thresholds along a spectrum: where should the model offer a statistical guess versus when should it admit uncertainty?
Users presumably want both answers and accuracy, but many times, optimizing for one requires limiting the other. Selecting a threshold along this spectrum is a necessary engineering tradeoff, not an attempt to mislead. Making these choices involves the same kind of “steering” the draft policy otherwise condemns.
Misconception #3: Bias mitigation makes models less accurate.
We’ve shown there is no “natural” state of an AI model, and no definitive “best” output. There are, however, ways to assess model performance against chosen standards to determine if it is useful its intended purpose. Among other things, this includes testing whether the model operates consistently across users and use cases – the exact type of training or assessment the Commission claims inversely impacts accuracy.
The definitive example is facial recognition. Early systems were trained on datasets composed primarily of white, male faces, but it turns out faces aren’t entirely generalizable: a model trained on a narrow subset works very well on that subset but decreasingly well on everyone else, such as women or those with darker skin. By one measure, those early systems were “accurate”; when matching white men, the technology worked well. But a system cannot be said to be “accurate” at scale when it performs poorly for whole categories of users. That is a demonstration of statistical bias: models whose performance varies based on a non-relevant factor. Thanks to improved training sets and “steering,” facial recognition systems now operate equally well, at an operationally high standard, across demographics.
The Commission, however, selects this sort of bias mitigation as its central example of accuracy suppression, describing it as the interposition of “so-called ‘equity’” at the expense of truth.
Whatever the Commission or the administration thinks of traditional DEI programs in other contexts, bias mitigation in AI models is not an ideological indulgence. Bias has a statistical and mathematical meaning and need not involve protected classes at all. It might be driven by geography, dialect, file formats, temporal attributes, or really anything at all. A model that advises women differently than men in response to the same question, or that gives degraded answers when a question is phrased in nonstandard English, is malfunctioning for those users. Neither the operator nor the user is well-served.
The effort required to first identify and then refine a model to mitigate these outcomes results in it being more accurate, not less. These practices represent risk-mitigation techniques recognized to reduce the likelihood of undesirable, or even illegal, performance outcomes. Liability under federal civil rights laws cannot just be pre-empted by the FTC’s reinterpretation of “reasonable expectations.”
An Unworkable Standard
The Commission’s proposed policy statement offers an unworkable regulatory standard. In demanding that AI systems produce accurate and neutral outputs, while failing to clarify what developers may or may not take into account, it directs enforcers to measure deviation from a benchmark that doesn’t exist and that technological realities do not allow.
The draft policy also willfully misapplies “reasonable expectations” – defining it in ways not consistent with past legal interpretation. It is just as likely that in a specific context, consumers may be assumed to expect warnings when errors may occur, to want a system to work safely, and to perform as well for them as for others.
The irony is that consumer protection law should not need a new benchmark. The Commission has long had a workable one in substantiation: What claims does a company make about its system, and can it back up those claims with evidence? Whatever performance metrics or representations does it apply, and has it tested them? Did it identify potential harms (including biased outcomes) and make efforts to mitigate them?
The FTC has already shown it will hold a provider accountable for overselling its systems or for misleading consumers. And frameworks like NIST’s AI RMF describe what this looks like in practice for AI systems specifically. This aligns with established consumer protection. There is no reason it cannot be sufficient for AI systems moving forward.