AI in the Boardroom: Why Directors' Safe Harbours May Not Hold

In May 2026, the Chief Justice of New South Wales, the Hon. Andrew Bell AC, delivered the 2026 Harold Ford Memorial Lecture at the University of Melbourne Law School. Its title tells you why it matters: “Corporate responsibility and directors’ duties in the era of Artificial Intelligence”. It is an address every director should read.

The Chief Justice was careful to put questions to directors rather than hand them answers, a deliberate and prudent stance for a sitting judge who may one day have to decide these very issues. Against what he called a “technological tsunami” with “no signs of abatement”, and with agentic AI “already upon us”, he set out to map the questions boards now face. They range from the provocative (could an AI bot itself be a “shadow director”?) to the immediate (do the Corporations Act’s familiar protections still hold when a board decision has been shaped by a machine?).

The questions he put to directors are pointed. Should AI literacy sit on the board’s skill matrix? How much must a director understand about the tools the company relies on? What paper trail should a board keep of the prompts behind an AI-assisted decision? Can a director even delegate a task to AI, and if they lean on its output, are they still protected? Most boards, he suggested, do not yet have good answers.

What he was prepared to call settled is more uncomfortable still. “Regulatory lag is inevitable,” he observed, but for directors, “lag is not really an option.” Several of the protections directors have long assumed they could fall back on may not be available once AI has played a prominent role in a board decision.

The safe harbours may not hold

Australian directors have three principal statutory protections to point to when a decision is later challenged: reasonable reliance under section 189 of the Corporations Act, the safe harbour for delegated acts under section 190, and the business judgment rule under section 180(2). The Chief Justice examined each, and in each case found that AI involvement strains the protection, sometimes to breaking point.

Reasonable reliance (s 189)

Section 189 allows a director to rely in good faith on information or advice prepared by an employee, a professional adviser, an expert or a fellow director, provided the director makes an independent assessment of it. The wording assumes the information was “given or prepared by” a person; an AI system is not one, so its raw output falls outside the section.

More fundamentally, the requirement to make an “independent assessment” runs straight into what the Chief Justice called the problem of “the black box”. How, he asked, can a director assess advice independently when neither the data behind it nor the reasoning that produced it is known or knowable? And where the advice comes from an adviser who leaned on AI, the director’s reasonable belief in that adviser’s competence may have to stretch to cover their responsible use of AI as well, an expectation no statute or decided case has yet mapped out.

Delegation (s 190)

Directors can delegate their powers under section 198D, but only to a committee, a fellow director, an employee or “any other person”. AI is not a “person” for the Act’s purposes. As the Chief Justice put it, “there is no power to delegate to AI”, and section 190’s protection does not apply to what is, in substance, a delegation to the AI itself.

The harder question, which he raised but did not finally resolve, is what happens when a properly appointed delegate performs a legitimately delegated task using AI. Does that become, in substance, an impermissible delegation to the AI under section 198D, for which section 190 offers no protection? Either way, a director seeking the section 190(2) safe harbour would still need to justify why their trust in that delegate was reasonable, given the known risks of sycophantic agreement, bias and hallucination. The arrival of agentic AI, systems that act autonomously towards a goal with limited supervision, only sharpens the problem.

The business judgment rule (s 180(2))

The business judgment rule protects a director who makes a decision in good faith, without a material personal interest, on an appropriately informed basis, and with a rational belief that it serves the company’s best interests. The Chief Justice’s conclusion here was the most pointed of all: “It must be extremely doubtful that a director who simply adopts the recommendation of an AI bot could rely on the business judgment rule defence.”

The reasoning draws on existing authority. A rational belief, the courts have held, requires “an arguable reasoning process”, and the rationality must be the director’s. The defence also requires a decision “consciously made so that judgment has actually been exercised”. A director who “blindly adopts” an AI recommendation has neither an arguable reasoning process they can articulate nor, arguably, made a conscious decision at all. For that director, the black box may well defeat the defence; even short of blind adoption, the Chief Justice noted, a lack of transparency in AI output “may impair or affect” a director’s ability to rely on the rule.

The paradox: ignoring AI may also expose you

The same provision may cut the other way. The Chief Justice noted a suggestion in the literature that, in some circumstances, the business judgment rule could require directors to use AI.

The rule asks directors to inform themselves “to the extent they reasonably believe to be appropriate”. As AI puts more (and better) information within reach, the bar for what counts as “appropriately informed” may rise with it. On that logic, a board that passed up an AI tool well suited to a data-heavy decision might one day have to explain why it did not.

Directors are squeezed from both sides, then. Lean on AI uncritically and the defence falls away; ignore it entirely and the same defence may weaken for the opposite reason. The lecture pointed to a sensible middle path: AI should be used “as an aide to [a director’s] own decision-making, rather than as a substitute for making an independent assessment”. The director must understand what the tool produced, be able to test it, and bring genuine judgment to bear. That is a higher bar than most boards currently set.

Shadow AI: the risk that is already here

If the safe harbour analysis is the longer-term concern, “shadow AI” is the one playing out in boardrooms right now.

Many boards’ first reaction is to clamp down, with approvals, tool restrictions and limits on what directors and executives may use. Well meant, the Chief Justice cautioned, but it can backfire. In his words, “frustrated directors or officers, without access to secure and sanctioned AI platforms offering embedded controls, may revert to ‘shadow’ AI use” on personal accounts and other consumer tools the organisation cannot see, secure or govern.

It is shadow IT all over again, except the asset now leaking out is the board’s own confidential material. The Chief Justice described shadow AI use as “apparently already widespread”, and the judicial observation behind that concern comes from ASIC v Bekier, the Star Entertainment governance judgment handed down in March 2026. That case was not about AI: Justice Lee found Star’s former chief executive and former chief legal and risk officer had breached their section 180(1) duties of care and diligence in relation to the company’s handling and escalation of serious anti-money-laundering and junket-related risks, while the seven non-executive directors were cleared. Board information overload was a theme of the judgment, not the ground of liability. But in what the Chief Justice describes as “the first Australian judgment to reflect on AI use by directors”, Justice Lee observed that “many individual directors are using AI informally to prepare for meetings” even though collective, board-endorsed use remains “far from entrenched” due to confidentiality and legal concerns.

The risks are concrete. A director feeding details of a confidential acquisition into a public AI tool has likely engaged the company’s confidentiality obligations and bypassed whatever data governance framework exists. And there is a forensic sting in the tail: it is “readily foreseeable”, the Chief Justice warned, that individual director AI use will trigger “a new avalanche of discovery” in regulatory inquiries, class actions and litigation. Prompts, transcripts and outputs may be discoverable, and used in cross-examination.

The cautionary tale he cited is the Delaware case Fortis Advisors v Krafton, in which a CEO used ChatGPT to shape and execute a strategy, largely followed its advice against his own lawyers’ warnings, and then deleted the chat logs. The AI-generated outputs were not privileged and were used as evidence against him at trial. The Chief Justice cited the episode to illustrate AI’s well-documented tendency towards sycophancy, affirming a user’s beliefs and preferences more readily than a human adviser would.

What the Chief Justice said is clear

Amid the open questions, the lecture set out several matters the Chief Justice described as settled:

• The framework is not changing, yet. For now, there is no move in Australia to alter the Corporations Act in response to AI. Existing duties apply, unmodified.
• The safe harbours may not be available. Where AI has played a prominent role in a decision and been relied upon by directors or their delegates, sections 189, 190 and 180(2) may not protect them.
• Directors must be vigilant and disciplined. Beyond confidentiality and data concerns, independence sits at the heart of good governance. Great care is needed to ensure “the crucible of debate and productive conflict of ideas in the board room” is not supplanted by AI or dulled into groupthink.
• Directors need genuine AI literacy. They must understand the limitations and risks of AI, both their own use of it and the company’s operational use, and have systems to mitigate those risks. The trend towards board AI education seminars is “salutary”.
• Balance matters. Locking down access without offering a secure, sanctioned alternative tends to hide AI use rather than stop it. The Chief Justice pointed to internal controls, visibility mechanisms and guardrails as a more realistic answer than an outright ban.
• A wave of litigation is foreseeable. It is “not difficult to imagine a wave of litigation” over the use, non-use and abuse of AI, with the bitter irony that AI may itself help formulate those very claims.

What your board should do now

The lecture is an Australian judgment-in-waiting, not a checklist. But the practical implications for any board are reasonably clear:

1. Treat AI literacy as a board capability. The Chief Justice openly asked whether AI competency should sit on the board’s skill matrix. An annual briefing over dinner will not cut it. Directors need a working grasp of where AI output can be trusted, where it cannot, and what to ask management about the company’s operational use of it.
2. Govern AI formally, not informally. Adopt clear policies rather than, in the words quoted in Bekier, “wink at” informal shadow use. Require management to disclose when AI was used to prepare board papers, which tools were used, how outputs were verified, and what the known limitations are.
3. Provide secure, sanctioned tools. If you want directors to stop using personal accounts on confidential matters, give them a secure, sanctioned platform with proper controls, and train people to use it.
4. Protect the integrity of the decision. For any significant decision where AI has been involved, each director should be able to articulate their own reasoning, independent of the AI output. If they cannot, the business judgment rule may not be available.
5. Mind the paper trail. Decide, in advance, what happens to prompts, transcripts and AI-drafted minutes, and remember that whatever is created may one day be discoverable.

The responsible use of AI, the Chief Justice concluded, “will fall to company directors and officers to deliver”. The frameworks already exist; the discipline does not yet. Boards that build that discipline now, through clear policies, real literacy, secure tools and genuinely independent judgment, will be far better placed than those waiting for the law to tell them what they already need to do.

This article summarises and comments on a public lecture and is general in nature. It is not legal advice. Directors should seek advice tailored to their own circumstances.