The Feds Say Chatbots Must Tell the Truth. Deciding What's True Is the Catch.
Washington wants to stop chatbots from lying to you — a goal I've argued for. But the same plan lets one agency decide what 'truthful' means. Here's the real trade.
Here is a sentence I did not expect to type in 2026: I think the Federal Trade Commission is onto something.
For years — in this newsletter, at dinner tables, to anyone who would sit still long enough — I have argued that a chatbot which quietly shades its answers to serve someone else's agenda is not a helpful assistant. It is a con. If a machine sells itself as neutral and is secretly steering you, that is deception in the plainest, most old-fashioned sense: a butcher with his thumb on the scale, charging you for meat that isn't there. Hold the machines to accuracy. Make them show the thumb. I have wanted precisely this for a long time.
So when the Commission announced it was going after chatbots that mislead the public, part of me wanted to stand up and applaud.
Then I read the next sentence. And the one after that. And the applause had to wait — because folded inside a consumer-protection announcement I mostly agree with is a lever I find genuinely frightening. The question is not whether AI should tell the truth. Of course it should. The question is who decides what "truthful" means — and in this proposal, the answer is a federal agency. Whoever holds that pen ends up owning the worldview of every machine that has to write to it.
Let me show you the thumb on this particular scale.
What the FTC actually did — and the one sentence that turns it into a weapon
On July 1, 2026, the Commission approved, on a 2-0 vote, a proposed policy statement on AI accuracy and began seeking public comment until July 31. Hold onto that word — proposed. This is not law yet. It is a draft the agency is floating, and the comment window is short. Chairman Andrew Ferguson framed the effort as protecting the public from what he called the subversion of AI systems for ideological ends.
The legal theory is elegant, and I mean that as a compliment before it becomes a warning. Section 5 of the FTC Act bans "unfair or deceptive acts or practices." The statement's argument is that if an AI company quietly steers a model's output for undisclosed ideological reasons, it is deceiving consumers — because consumers, the Commission writes, "have no basis to believe" the answers they get are distorted by hidden objectives. To make the point sting, the statement quotes the industry's own marketing back at it: Anthropic selling Claude as "a conversation with a brilliant friend," xAI pitching Grok as "your truth-seeking AI companion for unfiltered answers." A brilliant friend does not lie to you on someone else's instructions. And since roughly nine in ten users, by one cited figure, accept AI answers without fact-checking them, the deception lands on people with their guard down.
So far, this is a case I'd happily sign. Here is where it changes.
The statement did not arise on its own. It was ordered up by a December executive order that directed the government to hunt down state laws requiring "AI models to alter their truthful outputs." Read that phrase slowly. Truthful outputs. It presumes there is a settled, government-knowable truth, and that any state pushing a model away from it is the villain. And the statement names a villain by example: Colorado's recently revised AI Act. A company might steer its model, the statement allows, "including attempted compliance with a state law, such as Colorado's" — but doing so "may deceive consumers in violation of Section 5."
Then comes the sentence that turns a consumer-protection memo into a preemption weapon. The FTC Act, the statement concedes, "does not expressly preempt state law." But — and this is the whole ballgame — state law is impliedly preempted "to the extent it conflicts with a federal regulatory scheme." Translation: a state law that (in the agency's reading) forces an AI firm to shade its answers is said to collide with the federal duty not to deceive — and therefore, the theory goes, that state law loses.
To be clear about how load-bearing that move is: it rests entirely on implied conflict preemption — the courtroom doctrine that a state law falls when it can't coexist with a federal one, propped up here on a 1988 pipeline case. No statute says the FTC gets to erase state AI rules. This is the agency asserting it can, and daring a court to disagree. It has not been tested. It may not hold. But asserted loudly enough, in a live rulemaking, an untested theory still reshapes what companies dare to comply with.
Two things in the draft genuinely reassure me, and I want to be fair about them. First, honest mistakes are spared: the Commission says outright that hallucinations — a model confidently making things up — do "not in and of themselves" raise a Section 5 problem. The target is deliberate, hidden steering, not error. Second, there is an escape hatch: a company can steer, so long as it discloses — though the statement warns that an adequate disclaimer "could not be buried in terms of service." Disclosure over suppression. On its own terms, that is the right instinct.
So maybe this is good, actually?
Let me argue the other side as honestly as I can, because it is stronger than the reflexive objection.
Start with the uncomfortable fact at the center of all this. The single sharpest analysis I've read comes from legal scholar Alan Rozenshtein, writing at Lawfare, who concludes the administration's underlying "woke AI" order is likely constitutional as an exercise of the government's power over its own procurement. And he concedes the thing the critics keep dodging: thoroughgoing ideological neutrality in a model is a fantasy. There is no neutral "view from nowhere." Every model is trained on choices — what to include, what to downrank, what to refuse. "Unbiased AI" is not a real product. It is a marketing claim, and marketing claims are exactly what the FTC is allowed to police.
Follow that logic and the statement looks almost modest. If a model is being tuned toward some worldview no matter what, then a rule that says fine, tune it — just don't lie to the customer about it is not censorship. It is truth-in-labeling for software. I've demanded truth-in-labeling for everything from mortgages to breakfast cereal. Why would I flinch here?
Here is why I flinch. Rozenshtein's own verdict comes with a condition attached, and the condition is everything: all of it depends on the government "actually abiding by" the neutral principle "rather than simply using procurement policy as a cudgel to punish AI companies deemed too left-wing." That is the hinge. A power to define "truthful output" is safe only in the hands of an actor who never reaches for it as a bludgeon — and no such actor has ever existed for very long. The rule is only as honest as whoever is holding it this year. Which brings us to the part of the map Washington would rather you not look at.
Meanwhile, in the statehouses
While one federal agency reaches for the content of what a model says, the states have spent two years building rules about process — and the contrast tells you who is actually trying to govern this technology.
By the count of Tech Policy Press, by the midpoint of 2026 states had enacted 109 AI laws (plus 28 more on data centers). Overwhelmingly, these are not laws about what a chatbot may believe. They are laws about where automated systems make decisions that change your life. Colorado is the cleanest example — the very law the FTC holds up as its problem child. After a bruising fight (one HAIA has chronicled: Colorado wrote America's toughest AI law, then quietly repealed it before it could bite), the state repealed and reenacted the statute as SB 26-189 — a framework governing "consequential decisions" in education, employment, housing, lending, insurance, health care, and government benefits. It is, at bottom, a disclosure-and-documentation regime for the moments a machine decides whether you get the apartment, the loan, or the job.
Notice what that is not. It is not a rule telling a chatbot which side of a culture-war question to take. It is a rule saying: if an algorithm is deciding your mortgage, someone has to be able to explain it. That is process regulation — the same species of consumer protection the FTC claims to champion. And yet it is this law the federal statement names as the kind of thing that might be "impliedly preempted."
One more distinction the noise tends to flatten, because it matters. The lawsuit you may have read about — where the Trump Justice Department filed a complaint in intervention declaring that "embedding AI with state-mandated discrimination is a recipe for disaster" — targets Colorado's original 2024 law, not the revised one. And that original law was suspended before it ever took effect: a court stayed it, and the legislature replaced it. So the federal government is fighting the old Colorado law in court, while the new FTC statement takes aim at the watered-down replacement — the disclosure regime that survived. When even the compromise is declared too coercive, you learn that the objection was never really to the law's severity. It was to the state having a say at all.
Now run the tape forward
Here is the part that keeps me up, and it has nothing to do with which party wrote the December order.
Suppose the theory holds. Suppose a court agrees that one federal agency, applying its deception power, may define what counts as a model's "truthful output" and knock down any state rule that conflicts. You have now built a machine. And machines outlast the hands that build them.
Imagine it is 2031. A different administration, a differently composed Commission, a different account of what "truthful" and "distorted by ideology" mean. The whistle has changed hands. The same power that today polices steering in one direction now, just as lawfully, requires it in the other — because the standard was never "the truth," which no agency can hand you. It was "the truth as the current referee defines it." Every model in the country is retrained to the new definition. Four years later, an election, and it all flips again. The country's AI does not converge on reality; it oscillates on a federal timer, and the labs — rational, risk-averse, wanting to sell to the government — tune their systems to whichever way the weathervane is pointing.
That is the future implied by handing any single office the pen. Not a lie today. A permanent power to define the truth, waiting for whoever wins next.
What the smart people across the spectrum are saying
The reason I'm not writing this as a partisan gripe is that the alarm is bipartisan, and the people pulling it do not usually agree on anything.
From the free-market right, R Street warns that letting the federal government decide what constitutes "ideological bias" — even for something as limited as who gets a contract — "could set a dangerous precedent." That is a friendly-fire critique, aimed at an administration R Street is often aligned with. From the civil-liberties corner, the free-speech group FIRE, which has fought efforts to let the government decide what AI systems can say, calls government overreach "the most chilling threat" in the entire AI landscape. From the progressive side, Public Knowledge makes the mirror-image point: tearing down state AI laws while building no federal protections in their place is an abdication of leadership, not governance. And a straight-news account from Straight Arrow News puts the through-line plainly: wrapping this in consumer protection places the federal government "in a position to define truth and bias in the AI age."
Right, left, libertarian, and the reporters in the middle, all circling the same worry from different doors. That is usually a sign the worry is about the structure, not the party.
And the FTC statement is not acting alone. It is the agency-side twin of a move already underway in Congress, where a bipartisan discussion draft from Representatives Obernolte and Trahan would preempt state laws "specifically regulating the development" of AI models outright, with a three-year sunset. HAIA walked through that legislative bargain — a federal rulebook in exchange for silencing the states — a few weeks ago. What's new here is that you no longer need Congress to pass anything. An agency policy statement, if the courts bless its preemption theory, does much of the same work without a single floor vote.
What does this mean for you?
This one is unusually actionable, because the door is literally still open. Here is what I'd do:
Comment before July 31. The statement is a proposal, and the FTC has to read what comes in. You do not need to be a lawyer. A plain paragraph — I want AI held to accuracy, and I do not want any single agency to become the national arbiter of what "truthful" means — is a legitimate, on-point comment. Public rulemakings are one of the few places a normal citizen's words are entered into a record that a court can later see.
Separate the two questions in your own head, and demand your representatives do too. "Should chatbots be honest?" (yes) and "Should one federal office define honesty for all of them?" (very different question) are not the same fight. The whole rhetorical trick here is to make you answer the first and think you've answered the second.
Distrust the "brilliant friend" framing — from every direction. A model marketed as your unfiltered, truth-seeking companion is still a product with a worldview baked in by someone. That was true before the FTC weighed in and it will be true after. Assume steering exists; look for the disclosure; treat its absence as the tell.
Watch who holds the pen, not just what they write with it this year. The safeguard that matters is not whether you like the current definition of "truthful." It's whether you'd hand that same power to the people you trust least — because sooner or later, that is exactly who inherits it.
Keep an eye on the calendar. By the time you read this the comment window may have closed or a final version may have issued. If the "open for comment" framing is stale, the underlying fight is not.
My vote?
Hold AI to the truth. I mean that as sincerely as I've meant anything I've written here — a machine that cons you is a machine that should answer for it, and I'm glad someone with subpoena power finally said so out loud.
But "hold AI to the truth" and "let one office define the truth" are two different sentences wearing the same coat, and the entire fight is over which one you think you're endorsing. Accuracy is a floor we should all stand on. A federal monopoly on the meaning of "accurate" is a ceiling that comes down slowly, and never in the direction you expect. The honest version of this policy is the disclosure version — show the thumb on the scale — not the preemption version that quietly appoints a referee and hands the whistle to whoever wins the next election.
Truth is not something a regulator can issue like a license. The best we can do is force the machines to stop pretending they don't have a point of view, and then keep the power to describe reality spread across as many hands — states, courts, companies, and stubborn readers like you — as we possibly can. Concentrate that power in one place "for our protection," and we will have built the most persuasive propaganda engine in human history and helpfully labeled it Truth.
The HAIA Foundation exists for exactly this: keeping AI aligned to human autonomy instead of to whoever holds the loudest microphone. If that's a project you believe in, subscribe at substack.haia.foundation and follow the work at haia.foundation — the next front in this fight is always closer than it looks.