Omnilys

AI Disclosure

Last Updated: 2026-05-07
Version: 1.0

This page explains, in plain language, how Omnilys uses AI, what its limitations are, and what we expect from you when you use the output. It is referenced from our Terms of Service §9 and is incorporated by reference.

If you take only one thing from this page: the Service produces AI output, AI output can be wrong, you must verify it before acting on it, and you must not use it as the sole basis for decisions about people.

1. What the Service does

When you upload data and ask a question, Omnilys:

computes a summary of your data (statistics, sample rows, correlations);
sends the summary and your question to four large language models: Anthropic's Claude, OpenAI's GPT-4o, Google's Gemini, and xAI's Grok;
each model independently produces findings (observations, recommendations, business impact);
our consensus engine clusters the findings, scores how strongly the models agree, flags hallucinations and conflicts, and produces a unified set of "consensus findings";
a final report-writer model drafts a narrative report from the consensus output.

Every step in this pipeline is performed by AI. There is no human in the loop reviewing your output.

2. What "consensus" means and what it doesn't

Our consensus engine reduces the chance that a single model's hallucination ends up in your report. It does this by:

requiring multiple models to corroborate a finding before it is shown as high-confidence;
typing conflicts (direction, magnitude, severity) so disagreements are visible to you;
scoring each finding on whether it is "grounded" (cites numbers that actually appear in your data);
detecting "restatements" (where models all just quote the same line from the data summary, which is not the same as four independent analyses agreeing); and
flagging solo high-confidence findings from low-trust models as potential hallucinations.

What consensus does not do:

It does not guarantee any finding is correct. Four models can be wrong in the same direction, especially on data that resembles their training data.
It does not know your business. The models do not know what numbers are normal in your industry, what your strategic priorities are, or what trade-offs you are willing to accept.
It does not perform statistical hypothesis testing, regression analysis, or causal inference. Patterns it surfaces are descriptive, not causal.
It does not promise a specific accuracy rate. We make no claim that the Service is "X% accurate". Accuracy depends entirely on your data, your question, and the difficulty of the task.

3. Limits you must respect

These rules also appear in the Terms of Service. They are restated here so they are easy to find.

3.1 Findings are not professional advice

Findings are not financial advice, legal advice, tax advice, medical advice, employment advice, investment advice, or accounting advice. If your decision needs professional advice, get it from a qualified human professional.

3.2 You must verify before acting

Before you act on any specific number cited in a finding, independently verify that number against your source data. Do not assume the AI got it right.

3.3 Do not use Output as the sole basis for decisions about people

This is a hard rule. Do not use Omnilys output as the only or primary input into:

hiring, firing, promotion, compensation, or any other employment decision;
approving, denying, or pricing credit, lending, or insurance;
evaluating tenants, leases, or housing applications;
educational admissions or scoring;
medical diagnosis, treatment plans, or eligibility for medical services;
access to public benefits, immigration decisions, or law-enforcement decisions.

If Omnilys output ever informs a decision in any of those areas, a qualified human must independently review and confirm any conclusion before action is taken. This requirement reflects standard responsible-AI practice and applicable laws including (where relevant) the Colorado AI Act, Illinois AI rules, NYC Local Law 144, and the proposed federal frameworks under FTC and EEOC guidance.

If you are uncertain whether your use case falls into one of these categories, do not use the Service for that decision until you have confirmed it does not.

3.4 Do not use Output to mislead third parties

Do not present AI-generated findings to third parties (clients, investors, regulators, the public) without disclosing that they are AI-generated. Misrepresenting AI output as human professional analysis can violate fraud laws, consumer-protection laws, and professional-conduct rules.

3.5 Do not upload regulated data without the right contract

Do not upload data subject to HIPAA, GLBA, PCI-DSS, FERPA, or the GDPR's special categories (Article 9, including health, ethnicity, religion, biometrics, sexuality, political views) unless we have a signed agreement covering that use. The Service is not designed or contracted for these.

4. How to read a finding card

Every finding shown by Omnilys carries signals you can use to decide how much to trust it:

Signal	Meaning
Severity (high / medium / low)	The model's assessment of business impact. Always verify.
Confidence (e.g. 87%)	The consensus engine's combined confidence after weighing trust × agreement × diversity. Higher is better but not a guarantee.
Agreed by (e.g. "claude, gpt4o, gemini")	Which models corroborated the finding. More agreement = stronger evidence.
Multi-lab agreement badge	The corroborating models come from different AI labs. Independent labs hallucinate differently, so cross-lab agreement is stronger evidence than agreement among models from the same lab.
Grounded badge	The numbers cited in the finding match numbers found in your actual data. Higher grounding = the finding is consistent with the underlying data.
Verify badge	The finding has signals that suggest possible hallucination. Treat it as a hypothesis, not a fact.
Restated fact badge	All models just quoted the same line from the data summary; this is not "four AIs independently concluded X."
Conflicts tab	Where models disagreed. Read these. Sometimes the disagreement is more useful than the consensus.

5. What we do behind the scenes (transparency)

The trust score for each model adjusts over time based on objective per-run signals (agreement rate, hallucination rate, grounding rate, contradiction rate). Models that perform consistently get a higher voice in consensus.
The consensus engine's thresholds (when to merge two findings, when to flag a hallucination, when to award a multi-lab bonus) are documented in our codebase and are tuned over time. Current values can change without notice.
The report writer (the model that drafts the narrative) sees the consensus output, not raw model outputs. It is instructed to cite reconciled values and ranges, and to use language that signals certainty appropriately.
We log per-run "tuning stats" (statistical distributions, no Customer Data) to improve the consensus engine. These statistics are aggregated across users.

6. What you can ask us

"How was this finding produced?" — open a finding's "why?" link. The breakdown shows the math behind the confidence number.
"Which model said what?" — the Model Breakdown tab shows each AI's independent summary and finding count.
"Where did the numbers come from?" — source_columns on each finding shows which columns of your data fed into it.
"Can I see the raw model outputs?" — Pro users can request this via privacy@omnilys.com. Free users can re-run the analysis to see fresh outputs.

7. Reporting AI errors

If a finding looks fabricated, biased, or harmful, please flag it. Use the "Report inaccuracy" link on the finding card. Reports help us tune the consensus engine and the model-trust scores. Aggregated complaint patterns are taken seriously.

8. Updates to this disclosure

If we change how we use AI in a way that materially affects what you should expect from the Service, we will update this page and notify you.

Contact: ai@omnilys.com for questions about how the Service uses AI.

[LAWYER REVIEW] particularly for §3.3, against the most recent guidance under the Colorado AI Act and any 2026 updates to FTC/EEOC AI guidance.