The Forensic Witness: why automated decisions need an independent record

In a courtroom, a witness is someone who testifies to what they observed. A forensic witness is a specialised type: an independent party whose job is to observe, record, and preserve evidence in a way that cannot be altered without detection. The power of the forensic witness lies in separation: the entity that executes an action is not the entity that records it. That separation is what makes the evidence trustworthy.

The same concept already exists in financial systems. The auditor who signs off on financial statements is a forensic witness. The notary who witnesses a contract is a forensic witness. The flight recorder is a forensic witness. In each case, the record is created by someone who does not benefit from altering it, and its integrity can be checked by an outsider without asking the creator.

In 2025, 88 percent of organisations report using AI in at least one business function, and 71 percent say they regularly use generative AI, according to McKinsey's global survey. Yet Accenture's research with Stanford University finds that only 14 percent of organisations have put responsible AI into practice, and none have reached full maturity. Deloitte's 2025 board survey of 695 directors and C-suite executives across 56 countries finds that 31 percent still say AI is not on the board agenda, and 66 percent admit their boards have "limited to no knowledge or experience" with AI. The gap between adoption and accountability is wide.

This article explains what a forensic witness is in the legal sense, why automated decisions need an equivalent, and how the three major regulatory jurisdictions - Australia, the European Union, and the United States - are converging on the same answer: evidence, not logs, not dashboards, not screenshots.

The forensic witness, explained

In a courtroom, a witness is someone who testifies to what they observed. A forensic witness is a specialised type: an independent party whose job is to observe, record, and preserve evidence in a way that cannot be altered without detection. The power of the forensic witness lies in separation: the entity that executes an action is not the entity that records it. That separation is what makes the evidence trustworthy.

The same concept already exists in financial systems. The auditor who signs off on financial statements is a forensic witness. The notary who witnesses a contract is a forensic witness. The flight recorder is a forensic witness. In each case, the record is created by someone who does not benefit from altering it, and its integrity can be checked by an outsider without asking the creator.

When AI makes decisions that affect people's lives - credit approvals, insurance pricing, fraud flags, recruitment screening, healthcare triage - those decisions need the same protection. Most organisations today operate without one. The decision happens inside a model, the surrounding logs are fragmented across vendors, and nothing is signed in a way an outsider can verify. When a regulator, auditor, or court asks for proof, the honest answer is usually a story.

Why internal logs are not enough

The most common response to the accountability gap is to point at internal logs. Every agent action is logged, every API call is timestamped. But logs record events, not evidence. They live inside the same production environment that generated the decision, so they can be edited, overwritten, or accessed by the same administrators.

As EY's work on AI assessments explains, confidence in AI systems is essential, but it cannot rest on self-reported logs. The Organisation for Economic Co-operation and Development (OECD) reports that the monthly average rate of adverse AI incidents grew almost twenty-fold - from 32 in November 2022 to 614 in January 2025. EY's AI Sentiment Index finds that 58 percent of citizens are concerned that organisations are failing to hold themselves accountable for negative uses of AI, and 52 percent are concerned that organisations are failing to comply with AI internal policies and regulatory requirements.

To a forensic auditor, internal logs are a point of failure, not a control. The entity that executes the decision must not be the entity that witnesses it.

What the evidence must prove

An independent forensic witness for automated decisions must satisfy the same tests that courts and auditors have applied to evidence for decades:

1. Authenticity - the record was produced by the organisation's own key, not the vendor's.
2. Integrity - the record cannot be edited after the fact without that edit being obvious.
3. Time - the decision is anchored to an independently witnessed moment.
4. Provenance - the model version, inputs, and routing outcome are captured.
5. Retention + erasure - content can be deleted on request without destroying the audit trail.
6. Verifiability - anyone can check the above offline, with no account.

If any one of these is missing, you do not have audit evidence - you have a story someone told you about what happened.

What regulators are converging on

The push for independent evidence is not theoretical. It is being written into law across three major jurisdictions.

Australia: APP 1.7-1.9

From 10 December 2026, the Privacy and Other Legislation Amendment Act 2024 brings automated-decision-making transparency into the Privacy Act 1988. APP entities must disclose in their privacy policy the kinds of personal information used in automated decision-making, the kinds of decisions made solely by automation, and meaningful information about the logic. The Office of the Australian Information Commissioner (OAIC) released its Issues Paper in May 2026 and intends to publish guidance by September 2026.

The OAIC's reading is broad. The term "computer program" encompasses AI, machine learning, and rule-based systems. The "decision" limb captures not only fully automated decisions but also decisions where a human decision-maker is substantially and directly guided by a computer program. The OAIC explicitly notes that a human in the loop does not necessarily take an arrangement outside the framework.

For regulated entities, the obligation attaches to the organisation that arranged for the automated decision - not necessarily the developer or host. This makes vendor diligence and contractual audit rights critical, because the entity that procured the system is the one that must disclose how it was used.

Beyond the Privacy Act, APRA's prudential standards reinforce the same logic. CPS 234 (information security) and CPG 235 (data management) require that information assets remain effective and intact. CPS 232 (outsourcing) and CPG 229 (third-party risk) mean that the record must survive operator compromise and remain verifiable against the organisation's own key.

European Union: the EU AI Act

The EU AI Act (Regulation (EU) 2024/1689) takes a more technical approach. Article 12 requires that high-risk AI systems be technically capable of automatically recording events (logs) over their lifetime. These logs must enable traceability for three purposes: identifying situations that may result in risk, facilitating post-market monitoring, and monitoring operation by deployers. For remote biometric identification systems, Article 12(3) sets a specific minimum: session timestamps, reference databases, matched input data, and the identities of result-verifying personnel.

Article 19 requires providers to keep logs under their control for a period appropriate to the intended purpose, while Article 26 requires deployers to keep automatically generated logs for at least six months unless other applicable law requires longer. The Digital Omnibus agreed in May 2026 deferred the compliance deadline for stand-alone Annex III systems to 2 December 2027, but the design requirement is pre-market: the system must have the technical capacity to generate the required logs automatically before it is placed on the market.

United States: audit, evidence, and the NIST AI RMF

In the US, the framework is less consolidated but no less demanding. The NIST AI Risk Management Framework (AI RMF 1.0, released January 2023) asks organisations to establish mechanisms for AI auditability, traceability of the development process, and logging of processes and outcomes. The NIST AI RMF Playbook explicitly asks: "Did your organization implement a risk management system to address risks involved in deploying the identified AI solution?" and "Did you ensure that the AI system can be audited by independent third parties?"

At the same time, the Advisory Committee on Evidence Rules discussed proposed amendments to the Federal Rules of Evidence in April 2024. The Committee is considering a new Rule 707 to regulate machine-generated evidence when it is offered without an expert witness. The Committee's note explains that the problem with machine-learning evidence is not authenticity but reliability - and that reliability standards should apply whether the evidence is presented by a human expert or directly by the machine.

The PCAOB amendments to auditing standards (effective for fiscal years beginning on or after December 15, 2025) clarify auditor responsibilities when technology-assisted analysis is used, emphasising the reliability of external information and the importance of controls over information technology. The SEC's AI Advisory Committee has similarly flagged that audit committees need to understand how AI is used in financial reporting and whether internal controls are designed, implemented, and operating as intended.

What auditors and boards are saying

The demand for evidence is not just regulatory. It is coming from the people who have to defend decisions to regulators, courts, and shareholders.

KPMG's survey of 1,800 companies across six industries and ten countries finds that nearly two-thirds of respondents expect auditors to conduct a more detailed review of the control environment in relation to their use of AI in financial reporting. Fifty-three percent foresee auditors carrying out an AI governance maturity assessment, and one-third expect to ask auditors to provide third-party attestation over the use of AI technology.

PwC's 2025 Responsible AI survey of 310 US business leaders finds that 58 percent say Responsible AI initiatives improve return on investment and organisational efficiency, and 55 percent say they enhance customer experience and drive innovation. Yet only 61 percent say their organisations are at the strategic or embedded stage. PwC's 2026 Annual Corporate Directors Survey finds that 71 percent of directors say AI is the board capability most in need of strengthening.

Deloitte's board survey shows that while 40 percent of respondents say AI has caused them to think differently about their boards' makeup, only 5 percent report that AI is incorporated into their business and operating plans. The message from the boardroom is clear: AI is a strategic priority, but the evidence infrastructure is missing.

The six properties of a forensic decision record

A decision record that satisfies the forensic-witness standard needs six properties. Together, they close the gap between claiming a decision was made correctly and proving it.

These are not design preferences. They are the minimum an auditor needs to answer the question: "Can you prove what your AI decided, and that the record has not been changed?"

How to build the forensic witness layer

You do not need to replace your models. The forensic witness sits beside them, quietly producing the evidence your governance team already owes the regulator.

The architecture mirrors financial controls: separation of powers. The operational layer runs the AI. A separate Forensic Witness layer keeps a cryptographically signed, externally stored record of what happened, when, and with which inputs. Once that record is signed and chained, it becomes permanent - incapable of being manipulated, patched, or overwritten by any administrator within your production stack.

In practice, that means:

1. Capture each decision as a signed record with model version, routing outcome, and governance fields - not raw personal data.
2. Chain every entry to the previous one with a SHA-256 hash so tampering is detectable.
3. Timestamp with an independent RFC 3161 authority so "when" is provable.
4. Verify offline on a schedule. A failed verification should be treated as a severity-1 control failure.
5. Retain and erase on policy, keeping signed erasure proofs so deletion is demonstrable without destroying auditability.

The good news is that this is not a research project. It is a discipline of producing, for every automated decision, a signed, chained, time-anchored, verifiable record your organisation controls.

The term "forensic witness" has a precise meaning in law: an independent record that cannot be altered without detection, verifiable by an outsider without trust in the recorder. That is exactly what automated decisions need.

The regulatory timeline is real. Australia's APP 1.7-1.9 takes effect on 10 December 2026. The EU AI Act's high-risk logging requirements are in force now. US audit standards and the NIST AI RMF are pushing the same direction. The professional services firms - McKinsey, Deloitte, Accenture, PwC, EY, and KPMG - are all saying the same thing: the gap between AI adoption and AI governance is the defining risk of the next three years.

Every flight has a black box. Every automated decision should have a forensic witness. The question is not whether your organisation needs one. It is whether you will have it before the auditor, the regulator, or the court asks for it.

For a practical implementation guide for GRC teams, see Audit-ready AI: how GRC teams prove every automated decision.

• Verify a sample receipt with no account.
• Read the receipt specification.
• Book a demo to map the Forensic Witness standard to your stack.