Who this is for: AI product teams, general counsel at AI companies, compliance officers at organizations deploying consumer-facing AI, and policy teams preparing comment submissions.
Comment period closed July 31, 2026. Final guidance under FTC review. This is a policy statement, not a rule. It does not create new legal obligations. It signals how the FTC intends to apply the existing Section 5 deception framework to AI systems. The message is clear: if you steer an AI model's output away from the most accurate answer without telling users, the FTC may treat that as deception.
Contents
1. What the Policy Statement Says 2. What This Means in Practice 3. Disclosure Requirements 4. SWT3 Procedure Cards 5. Comment Period and Next Steps 6. Framework Mapping 7. Quick Reference 8. Quick Start 9. References1. What the Policy Statement Says
The FTC's July 1, 2026 policy statement introduces the concept of "suppression of accuracy": deliberately training or configuring an AI model to prioritize objectives other than the most accurate answer, without clear and conspicuous disclosure to users.
This is not about AI hallucinations or unintentional errors. The statement targets deliberate output steering, where a company makes an intentional product decision to have the model produce something other than its most accurate output, without informing users that this is happening.
Examples of suppression the FTC identifies:
- Tuning a model to favor certain products or viewpoints in recommendations
- Filtering accurate but commercially inconvenient responses
- Training on datasets selected to produce skewed outputs aligned with business interests
- Configuring system prompts that direct the model away from accurate but unfavorable answers
The legal framework is Section 5 of the FTC Act, which prohibits "unfair or deceptive acts or practices." The FTC is applying the same deception test it has used for decades in advertising and consumer protection to a new product category. The framework is not new. The application is.
The disclosure standard is "clear and conspicuous." The statement explicitly identifies burying a disclaimer in terms of service as insufficient. A one-time disclosure tucked away in fine print is unlikely to satisfy the standard. Companies are allowed to prioritize objectives other than pure correctness, but only with transparent, ongoing disclosure.
2. What This Means in Practice
The practical implications depend on how your AI system is configured and deployed. Work through these questions:
Is your AI system configured to prefer certain answers over more accurate alternatives? If a recommendation engine, chatbot, or search assistant is tuned to surface specific products, services, or viewpoints, users must be told. This includes system-level configurations like temperature adjustments designed to reduce the probability of certain accurate but unwanted responses.
Do your guardrails suppress accurate information? Safety filters, content policies, and topic restrictions are not prohibited. The FTC does not require AI systems to answer every question fully. It requires that the existence and general nature of filtering be disclosed when that filtering affects accuracy in ways that are material to users.
Was your model fine-tuned on data that introduces systematic bias toward certain outcomes? If fine-tuning on proprietary data causes the model to systematically favor certain conclusions, that bias and its commercial or editorial source should be disclosed.
The critical nuance: the FTC is not requiring AI systems to be perfectly accurate. It is requiring transparency about what objectives the system optimizes for when those objectives differ from pure accuracy. A model optimized for engagement, for brand safety, or for commercial outcomes is not inherently problematic. An undisclosed optimization is.
The statement explicitly addresses advertising contexts: AI-powered recommendation systems, search assistants, and chatbots that steer users toward specific products or services are squarely within scope. If your AI system has a commercial interest in the outcomes it recommends, that interest must be visible to users.
3. Disclosure Requirements
| What the FTC Expects | What Is Insufficient | SWT3 Evidence |
|---|---|---|
| Clear statement of what objectives the AI optimizes for beyond accuracy | Terms of service buried disclaimer | AI-TRANS.1 records transparency disclosures per interaction |
| Conspicuous notice visible to users at point of interaction | One-time disclosure at signup only | AI-INF.1 timestamps prove ongoing disclosure at inference time |
| Specificity about the nature of output steering: commercial, editorial, or safety | Generic "AI may not always be accurate" disclaimer | AI-BASE.1 records model configuration including optimization objectives |
| Updated disclosure if steering parameters change after deployment | No disclosure update after model fine-tuning or configuration change | AI-BASE.1 configuration drift anchors show when parameters changed and when disclosure was updated |
4. SWT3 Procedure Cards
Five SWT3 procedures produce the evidence chain most relevant to FTC accuracy disclosure requirements.
Inference Provenance
FTC context: If a user complains that an AI system gave them biased or steered information, the first question is: what did the system actually output, and when? AI-INF.1 provides the cryptographic record of every inference, including the model version, the input hash, and the output hash.
How SWT3 witnesses it: Every inference produces a Witness Anchor. If the FTC investigates, the inference record shows exactly what the system produced and under what configuration. The anchor is immutable and timestamped, meaning the record cannot be revised after the fact.
AI-INF.1 anchors prove what the system said. Compare against the model's optimization objectives documented in AI-BASE.1 to demonstrate whether output steering was disclosed at the time of inference. Gaps in the inference chain during periods of active user interaction are a significant finding.
Transparency Record
FTC context: The policy statement's core requirement is disclosure. AI-TRANS.1 records what transparency measures were active during each interaction: what disclosures were shown to users, whether the system identified itself as AI, and what limitations were communicated.
How SWT3 witnesses it: Creates verifiable evidence that disclosure obligations were met at the point of interaction, not just configured in a policy document. The anchor records disclosure_type, disclosure_text, and whether the disclosure was conspicuous or buried.
AI-TRANS.1 anchors are the primary evidence for FTC compliance. They prove disclosure happened, when it happened, and what was disclosed. For FTC purposes, verify that disclosure_type reflects the specific nature of output steering (commercial, safety, editorial) rather than a generic accuracy disclaimer.
Guardrail Enforcement
FTC context: Guardrails that suppress accurate information, including content filters, safety classifiers, and topic restrictions, are a form of output steering. The FTC does not prohibit guardrails, but requires disclosure of their existence and nature when they affect accuracy in ways material to users.
How SWT3 witnesses it: AI-GRD.1 records guardrail status and whether guardrails triggered for each inference. Creates evidence of when guardrails affected output, what type of guardrail was applied, and whether the user was notified.
AI-GRD.1 anchors show which inferences were affected by guardrails. This supports the disclosure argument: the company had guardrails, told users about them, and the anchors prove they were active. Cross-reference with AI-TRANS.1 to verify that guardrail triggers were accompanied by a corresponding user-facing disclosure.
Baseline Configuration
FTC context: The model's configuration defines its output steering behavior. Temperature settings, system prompts, fine-tuning objectives, content policies, and optimization targets all affect what the model produces. AI-BASE.1 records these parameters at each point in time.
How SWT3 witnesses it: Configuration changes produce new anchors. If a model is re-tuned to favor certain outputs, the configuration change is recorded with a timestamp. This creates an auditable history of when steering parameters changed and, paired with AI-TRANS.1, when disclosure was updated to reflect those changes.
AI-BASE.1 anchors document the model's optimization objectives over time. Critical for demonstrating that disclosure was updated when steering parameters changed. Look for AI-TRANS.1 anchors with updated disclosure text within a reasonable window of each AI-BASE.1 configuration change. A configuration change with no corresponding disclosure update is a gap.
Explainability Evidence
FTC context: The FTC expects companies to be able to explain why their AI system produced a particular output. AI-EXPL.1 records explainability artifacts: feature importance scores, attention patterns, or reasoning traces that show how the model arrived at its output.
How SWT3 witnesses it: Records the explainability method used, the confidence score, and the key factors in the output. Creates evidence that the company can explain its AI system's behavior when a specific output is questioned.
AI-EXPL.1 anchors demonstrate good-faith effort to understand and explain AI outputs. Supporting evidence if the FTC questions why a system produced a particular response. Explainability records are also useful for distinguishing intentional steering (which requires disclosure) from unintentional model behavior (which requires different remediation).
5. Comment Period and Next Steps
The comment period closed July 31, 2026. The FTC is now reviewing submissions and developing final guidance. Organizations that submitted comments shaped the record; those that did not should prepare for the final statement as issued.
Key areas where the policy statement leaves room for clarification:
- Scope of "suppression": Does safety filtering that prevents accurate but harmful responses constitute suppression? The statement is not fully resolved on this point. Organizations with safety-critical AI systems should request explicit carveouts.
- Disclosure standard for AI: What is "clear and conspicuous" in a chat interface versus a voice assistant versus an embedded recommendation widget? The standard is borrowed from advertising law but the implementation details for AI are not specified.
- Enforcement timeline expectations: The statement is not a rule, but it signals investigation intent. Organizations should request clarity on whether the FTC expects retroactive disclosure for systems already deployed.
Regardless of whether the policy statement changes after comments, the direction is clear: the FTC views undisclosed output steering as deception. Begin documenting your AI system's optimization objectives, guardrail behavior, and disclosure practices now. The evidence chain is valuable regardless of the final policy, and organizations with documented evidence chains will be in a substantially better position if an investigation occurs.
6. Framework Mapping
| Framework | Relevant Requirements | SWT3 Evidence |
|---|---|---|
| FTC Section 5 | Deception, unfair practices, disclosure of output steering objectives | AI-TRANS.1, AI-BASE.1, AI-INF.1 |
| EU AI Act Art. 50 | Transparency obligations for AI systems interacting with natural persons | AI-TRANS.1, AI-MARK.1 |
| NIST AI RMF | GOVERN 1.1 accountability, MAP 1.5 risk identification, MEASURE 2.5 bias evaluation | AI-INF.1, AI-BASE.1, AI-EXPL.1 |
| California SB 942 | AI transparency disclosure, content detection for synthetic media | AI-TRANS.1, AI-WATERMARK.1 |
7. Quick Reference
| FTC Investigator Question | Where to Look |
|---|---|
| Does this AI system prioritize any objective other than accuracy? | AI-BASE.1 anchors with optimization_objectives field. Any objective beyond accuracy should be cross-referenced with AI-TRANS.1 to verify disclosure. |
| Were users told about output steering at the point of interaction? | AI-TRANS.1 anchors with disclosure_type and disclosure_text. Timestamps must align with or predate the corresponding AI-INF.1 inference records. One-time signup disclosure alone is insufficient. |
| Did guardrails affect the accuracy of outputs shown to users? | AI-GRD.1 anchors with guardrail_triggered = true. Cross-reference with AI-TRANS.1 to verify that user-facing disclosure accompanied the guardrail activation. |
| When was the model's steering configuration last changed? | AI-BASE.1 configuration drift anchors, sorted by timestamp. Each change creates a new anchor. Compare with prior anchors to identify parameter changes and verify disclosure was updated accordingly. |
| Can the company explain why a specific output was produced? | AI-EXPL.1 anchors linked to the inference in question via inference_id. Check explainability_method and key_factors fields for substantive explanation rather than placeholder values. |
| Was there any gap in the inference record during the period in question? | AI-INF.1 anchor continuity. Sort by timestamp and check for gaps. A gap during an active deployment period indicates either a logging failure or a period of unwitnessed operation. |
8. Quick Start
pip install swt3-ai
from swt3_ai import WitnessClient
client = WitnessClient(
tenant_id="your-tenant-id",
api_key="axm_live_..."
)
# Record model configuration and optimization objectives
client.witness_baseline(
model_id="your-model",
optimization_objectives=["engagement", "brand_safety"],
accuracy_priority="secondary"
)
# Record transparency disclosure at inference time
client.witness_transparency(
disclosure_type="commercial_steering",
disclosure_text="This assistant may favor products from our partners.",
conspicuous=True
)
# Record the inference with provenance
client.witness_inference(
model_id="your-model",
input_hash=sha256(user_input),
output_hash=sha256(model_output)
)
# Run the demo to see it in action
python -m swt3_ai.demo
Full SDK documentation: sovereign.tenova.io/docs
Create a free account: sovereign.tenova.io/signup
9. References
- FTC Policy Statement on Suppression of Accuracy in AI Systems (Federal Register, July 1, 2026)
- FTC Comment Portal (regulations.gov, comment period closed July 31, 2026)
- EU AI Act Transparency Obligations Guide (SWT3 Protocol)
- NIST AI RMF Compliance Guide (SWT3 Protocol)
- Cryptographic AI Evidence Quickstart (SWT3 Protocol)
- SWT3 SDK Documentation (sovereign.tenova.io/docs)
- Create a free account (sovereign.tenova.io/signup)