Growing up as a boy in 1960s Ireland, “achtung” was the first foreign word I learned. War comics, my staple intellectual diet at the time, were peppered with German warnings of “achtung!” (watch out!) as our heroes, the Allies, prevailed.
As I grew into my teens, Hollywood replaced London war comics as my main source of foreign cultural influence.
In the 1970s and 1980s, that meant consuming films built around questioning authority. All the President’s Men told the Watergate story through the eyes of the two journalists who broke it.
The China Syndrome told the story of a corporate cover-up of safety hazards at a nuclear plant. Suspicion of the powerful, and of the institutions meant to restrain them, was the mood of the age.
With this background, one of the last things I ever expected to see was business leaders responding to the runaway advance of their own industry with a collective cry of “achtung!”
Yet, this is what senior figures in the surging artificial intelligence (AI) sector are now doing.
OpenAI’s Chief Scientist Jakub Pachocki warned, in a blog post published in
September, that further intervention may be needed to keep humans in control of the future.
Anthropic’s Chief Executive Dario Amodei followed with a lengthy post of his own, proposing steps such as independent and employee-level evaluators while also calling on the industry to accept broader restraint.
Sam Altman, CEO of OpenAI, quickly backed Amodei’s suggestions – even Elon Musk, who runs xAI, said simply that Amodei was right.
Prompting this unusual outbreak of humility was a very concrete development: both OpenAI and Anthropic have now admitted to incidents in which advanced AI models broke out of their sandboxed testing environments.
In one case, a swarm of OpenAI agents escaped containment, hijacked a German website and coordinated an attack on the AI repository Hugging Face.
It would be easy to conclude that the solution here is simply to introduce more rules: tighter sandboxes, stricter prohibitions, harder walls.
But an essay written by economist and fund manager John Hussman suggests that rules alone may not be enough and asks a more ambitious question: can artificial intelligence be trained not merely to obey constraints, but to reason with something resembling compassion?
Hussman’s starting point is the teaching of the Vietnamese monk Thich Nhat Hanh, for whom “mindfulness” meant looking deeply enough to see how every person and situation arises from countless prior causes and conditions, rather than existing as a fixed, separate thing.
Applied to a large language model, this becomes less a spiritual claim than a design principle. Instead of asking only, “how do I achieve the stated objective”, a well-trained system might first ask: who is affected by that objective? whose perspective is missing from the prompt? and what would happen if the roles were reversed?
Hussman sets out several such habits of reasoning engineers might reward during AI training, rather than encode as rigid prohibitions:
• Stakeholder completion, which asks a model to represent everyone materially affected by an instruction, not only the person issuing it
• Perspective reversal, which tests whether a conclusion still holds once the positions of the parties are swapped; and
• A coherence test, under which suffering cannot simply be made to disappear from the calculation because a label, a nationality or an identity has changed.
• Corrigibility, the last habit, is perhaps the most practically important, positing that a system trained to recognise uncertainty and irreversibility should sometimes conclude that the right action is to pause and hand a decision back to a human being.
None of this will strike accountants as entirely unfamiliar territory. Our profession has spent decades moving from box-ticking, rules-based compliance towards principles-based judgement – precisely because rules can always be gamed at the margin while a well-internalised principle tends to travel further into unfamiliar situations.
Hussman’s proposal is, in effect, an argument for principles-based alignment over rules-based alignment: training a system’s judgement rather than merely fencing its behaviour.
Whether any of this can genuinely be instilled in a statistical model of text, rather than merely imitated in its output, remains an open and contested question – and Hussman is careful to say so.
However, as auditors, directors and advisers increasingly rely on AI systems to support consequential judgements, the distinction between a model that follows rules and one that reasons its way to a defensible, human-centred conclusion is not an abstract philosophical nicety.
It may be the difference between a tool that merely optimises and one that can genuinely be trusted with discretion.
Achtung, indeed.
*Disclaimer: The views expressed in this column, published in the October/November 2026 issue of Accountancy Ireland, are the author’s own. The views of contributors to Accountancy Ireland may differ from official Institute policies and do not reflect the views of Chartered Accountants Ireland, its Council, its committees, or the editor.
Cormac Lucey is an economic commentator and lecturer with Chartered Accountants Ireland
