AI Ethics Commitment

Human judgment comes first

Sift Labs does not build AI to replace prosecutors, investigators, judges, defense counsel, or agency decision-makers.

Our AI systems may help legal professionals review and work with evidence across formats, including video, audio, documents, transcripts, images, and other case materials. Today, that may mean helping a user locate moments in footage, suggest edits, generate captions, summarize video, or prepare a working copy. As Sift Labs grows, the responsibility boundary stays the same: our AI does not decide who should be charged, whether a person is credible, what evidence must be disclosed, whether a case should resolve, or whether a person is guilty.

AI output must be reviewed by an authorized human before it is used in a case, shared outside the agency, or presented in court.

Appropriate use in criminal justice

Sift Labs products are tools for evidence workflow. They are not systems for determining guilt, dangerousness, credibility, charging, detention, plea terms, sentencing, or punishment.

Perry may help a user find a moment in a video, create a caption draft, prepare a clip, or generate a working copy. Those outputs still require case-aware legal review. AI assistance does not replace a prosecutor's obligations around accuracy, fairness, disclosure, candor, or professional judgment.

Evidence integrity

Original evidence should remain original.

Sift Evidence is designed so source evidence is preserved and edits happen on working copies. AI-assisted actions are treated as assistance, not as a substitute for evidentiary review.

• Generated outputs remain reviewable before export.
• Edits are attributable to users and recorded in an audit trail.
• We design for a defensible chain of custody, not for shortcuts around it.

For details on encryption, access controls, audit logging, and CJIS-aligned security practices, see our Security page.

AI data boundaries

AI systems are only as trustworthy as the data boundaries around them.

Sift Labs designs evidence-processing workflows so raw evidence video, audio, and frames are not casually sent to uncontrolled general-purpose AI tools. Any use of external services is reviewed against the sensitivity of the data, customer agreements, security requirements, and applicable criminal justice obligations.

Our commitments are:

• We do not sell customer evidence or case data.
• We do not use customer evidence for advertising.
• We do not use customer evidence to train public AI models.
• We do not expose raw evidence to third-party AI systems unless that use is authorized, reviewed, and governed by the applicable customer agreement.
• We design prompts, model inputs, and generated outputs with data minimization in mind.

Accuracy and verification

AI can be useful and wrong at the same time.

Sift Labs designs AI outputs to be reviewed, corrected, accepted, or rejected by users. We do not ask prosecutors to trust unreviewed AI output. Captions may need correction. Object detection may miss something. A summary may omit context. A suggested edit may not be appropriate for court.

That is why our products emphasize reviewable workflows, preserved originals, and user-controlled export.

Bias and limitations

AI systems can perform differently depending on video quality, lighting, camera angle, audio quality, accents, language, movement, occlusion, and other real-world conditions. They can also reflect limitations in training data, evaluation methods, and deployment context.

Sift Labs will not represent AI as neutral simply because it is automated. We design for testing, monitoring, human review, and clear product boundaries.

Auditability and accountability

Criminal justice AI must be accountable after the fact.

Sift Labs designs its products to preserve meaningful records of important actions: who acted, what changed, when it happened, and what output was created. Audit logs are intended to support agency oversight, internal review, security investigation, and evidence integrity.

We also design for explainable workflows. A prosecutor should be able to understand how a clip, caption, annotation, or export was produced without relying on a black-box assurance that "the AI did it."

What we do not do

Sift Labs does not:

• Replace prosecutor judgment.
• Decide guilt, credibility, charging, detention, plea terms, sentencing, or punishment.
• Modify original evidence as part of normal editing workflows.
• Treat AI output as verified evidence without human review.
• Use customer evidence for advertising.
• Sell customer evidence or case data.
• Claim AI is objective merely because it is automated.
• Ask agencies to accept vague AI claims in place of documentation.

Continuous review

AI ethics is not a one-time statement. It is an operating practice.

Sift Labs will continue reviewing its products, model behavior, vendors, customer feedback, court expectations, public-sector procurement standards, and criminal justice requirements as they evolve.

Guidance we track

We track public-sector and AI-risk guidance, including the NIST AI Risk Management Framework, Council on Criminal Justice AI principles, and OWASP Top 10 for LLM Applications.

Questions about this statement can be sent through our connect form or by mail to Sift Labs Corp., 21812 Oxford Dr, Lago Vista, TX 78645.