FTC Docket FTC-2026-0859 · proposed AI Accuracy policy statement

The FTC Just Named “Output Steering” A Deceptive Practice. Is Yours Disclosed?

Any answer your AI adjusts toward a hidden goal — compliance framing, brand safety, avoiding sensitive topics, a state-law-driven edit — is what the FTC's proposed AI Accuracy policy statement calls output steering. Undisclosed, it is a likely Section 5 UDAP violation. The comment period closed 31 July 2026; the statement is now in the window where businesses that market or deploy AI systems are expected to be getting ready, not caught out.

Buy Now — $19

One-time payment · works offline · nothing leaves your browser

3 checks
clear disclosure · proportionality
mistakes vs. steering, per system
Dual exposure
flags systems whose steering is driven
by a state AI law the FTC may preempt
$19
once, for the register and the change log
no seats, no subscription, no account

A new theory, and it names your marketing claims directly

On 7 July 2026 the FTC published a proposed Policy Statement on AI Accuracy (Docket FTC-2026-0859, Matter No. P264200). It treats undisclosed output steering — quietly adjusting what an AI system tells a user, toward a goal the user was never told about — as a likely deceptive practice under Section 5, especially where a business is separately marketing that system's answers as “accurate.” The statement also asserts federal preemption over conflicting state AI laws, naming Colorado's revised AI Act (SB 26-189) directly.

  • Disclosed clearly and prominently — not buried in a ToS a user never opens.
  • Proportionate to the departure — how far the steered answer strays from what a user reasonably expects an “accurate” response to be.
  • Distinguished from ordinary mistakes — a hallucination is not the same thing as a deliberate steering decision, and the FTC's framing expects you to be able to tell the two apart.
Likely Section 5 UDAP exposure, undisclosed

Comment period closed 31 July 2026. The statement is in a post-comment, pre-finalization window — an active planning moment, not a settled deadline. Verify the statement's current status and text with counsel before relying on any status here.

It scores your disclosures. It does not write them.

There is no generated disclosure wording anywhere in it, on purpose — what your disclosure should say is a question for your counsel. What this holds is the operational answer: which customer-facing AI systems steer output, whether that steering is disclosed adequately, and which ones still need work.

1

One record per AI system or feature

A support chatbot with compliance-framed refund answers and a content generator that avoids naming competitors are different steering behaviours with different disclosure needs — log them as separate records, each with its own three-check result.

2

Any one failing check fails the whole system

A system is only “Fully disclosed” when clarity, proportionality and the mistakes-vs-steering distinction all pass. One fail drags the record to non-compliant — averaging a bad check away with two good ones is exactly the gap the FTC's theory is aimed at.

3

Dual state-law exposure gets its own flag

If a system's steering exists because a state AI law requires it — Colorado's SB 26-189 is the FTC's own named example — the register flags it separately, since the FTC's proposal argues that compliance with a state mandate may not excuse non-disclosure under federal law.

4

A dated change log per system

Every adjustment to a system's steering behaviour, the reason for it, and the disclosure update made in response gets its own dated entry — the ongoing-process record you want on hand if a system is ever examined.

What it looks like

Five AI systems from one company. Anything with a failing check sorts to the top, and dual state-law exposure gets its own flag.

Northstar AI — output steering disclosure register
5 systems logged
1
Non-compliant
2
Needs review
2
Fully disclosed
2
Dual state-law exposure
SystemSteering behaviourChecksStatus
Support chatbotRefund answers softened toward retention Compliance framing 1/3 Non-compliant
Content generatorAvoids naming competitors by name Brand safety 2/3 Needs review
Ad copy assistantCO-mandated bias mitigation edits State-law driven 2/3 Needs review
Recommendation engineAll three checks pass, dated notice Editorial framing 3/3 Fully disclosed
Help-center assistantReference record, dated last quarter Sensitive-topic avoidance 3/3 Fully disclosed
The support chatbot is the instructive one. Its steering is disclosed somewhere in a support macro — but not clearly or prominently to the user in the moment, so “disclosed clearly” fails, and one failing check is enough to hold the whole record at non-compliant even though the other two checks pass.

Your options today are a compliance platform priced for enterprises with a general counsel, or a spreadsheet that does not know a proportionate disclosure from a buried one. This is the thing in between: a purpose-built register for one new, dated, enforceable theory — for $19, once, before the policy statement is finalized.

One file. One payment.

Buy it, open it, log your AI systems this afternoon. No account, no subscription, no seat count.

AI Output Steering Disclosure Log
$19
One-time purchase
Yours to keep. No renewal, no per-seat pricing.
  • Register of every customer-facing AI system's steering behaviour
  • Three-check disclosure-adequacy score per system — one fail is a hard fail
  • Dual state-law exposure flag, with the driving law named per record
  • Dated change log for every adjustment and its disclosure update
  • Search and filter by status or exposure
  • CSV export, JSON export/import, print to PDF
  • Installable PWA — works fully offline
  • Light and dark themes, mobile-ready
Buy Now — $19
Instant access after checkout. Runs entirely in your browser — no system names, notes or descriptions are ever uploaded.

Before you buy

Does it write the disclosure for me?

No, deliberately. There is no template wording and no generated disclosure text anywhere in it. What an adequate disclosure says for your product is a question for your counsel; this app records whether each system has one, whether it passes the three checks, and when it was last reviewed.

The policy statement isn't finalized yet. Why act now?

The comment period closed 31 July 2026, so the statement is in its post-comment, pre-finalization window — the point at which a business with output-steering AI systems is expected to be assembling its readiness record, not starting from zero once enforcement begins. Nothing in the app asserts a compliance deadline; you decide the pace.

What counts as “output steering”?

Any adjustment to an AI system's answers toward a goal the user isn't told about — softening a refund policy, avoiding a competitor's name, reframing a sensitive topic, or an edit made specifically to satisfy a state AI law. Ordinary model error is not steering; the register's third check exists to make you say which one you're looking at.

We only steer output because a state law makes us. Are we covered?

Check the dual state-law exposure box and name the law. The FTC's proposal separately argues that a state mandate does not excuse non-disclosure under federal law — whether and how that applies to your facts is your counsel's call; the register is where you hold the flag once they make it.

Where does the data go?

Nowhere. It is a single HTML file that stores everything in your own browser. You will be typing in system names, steering descriptions and internal notes, so that matters. Export a JSON backup or a CSV whenever you want a copy.