Anthropic has done something unusual for a frontier AI company: it has publicly drawn limits.
The US-based developer of advanced language models has signalled boundaries around certain military and surveillance applications of its technology, at a moment when artificial intelligence is rapidly embedding itself in national security, critical infrastructure and information ecosystems, and when the current US administration wants to deprioritise AI safety architecture in favour of “strategic advantage”.
In an industry defined by scale, speed and geopolitical rivalry, Anthropic’s restraint stands out. But an equally consequential question is not about Anthropic’s responsibilities. It is about the capital that backs tech companies and the investors behind it.
If companies building foundational AI systems now sit at the intersection of dual-use risk – national security policy and potential systemic human rights harm – then they are no longer just growth stocks. They are tests of governance, and of investor credibility.
And this exposes an increasingly convenient argument circulating in capital markets: that the largest technology firms are simply too “crucial” to divest.
According to this view, fund managers mandated to track global indices cannot realistically disengage from dominant technology companies, regardless of the severity of adverse human rights impacts or the limits of investor leverage. Big tech, the argument goes, is structurally indispensable to diversified portfolios.
Even large asset managers that publicly commit to integrating human rights due diligence in line with the UN Guiding Principles on Business and Human Rights (UNGP) and the OECD’s framework on responsible business conduct rely on this logic.
Mostly, the hook for this position comes from a passage in the OECD Due Diligence Guidance for Responsible Business Conduct for institutional investors:
“In some instances, the company may represent a crucial business relationship, or investment. Under the UN Guiding Principles, a relationship could be deemed as crucial if it provides a product or service that is essential to the enterprise’s business, and for which no reasonable alternative source exists.
“A relationship may be complex to unwind. That does not place it beyond scrutiny”
“In the context of investment, this is likely to be less relevant to specific investee companies but may be relevant for categories of companies of a given size or sector which are necessary to adequately diversify a portfolio (eg high market-cap energy companies).”
This language is increasingly used to suggest that divestment from large technology companies is effectively off the table, and that responsibility therefore begins and ends with engagement.
It is not an implausible reading. But it is a dangerous extension.
‘Crucial’ does not mean immune
Under the UNGPs, a “crucial relationship” does not extinguish responsibility. It alters sequencing.
Even where a relationship is considered crucial, investors are still expected to intensify leverage and seek to prevent and mitigate adverse impacts. Where progress stalls, engagement should be escalated and investors must assess whether continued involvement itself contributes to harm.
If mitigation efforts prove unsuccessful, divestment should be considered as a last resort – or investors must be prepared to accept potential reputational, financial or legal consequences. There is no safe harbour clause for systemically important companies.
The concept was originally conceived for situations such as dependence on a sole-source supplier of an essential input. Extending that logic into “the largest firms in the index are untouchable” stretches it beyond recognition.
A relationship may be complex to unwind. That does not place it beyond scrutiny.
Passive is not neutral
The index-tracking defence also obscures something else: mandates are design choices.
A manager may be required to hold index constituents, but this does not mean that weighting flexibility is absent, that alternative index products cannot be constructed, or that sector exclusions are legally impossible. Nor does it render escalation strategies futile.
Portfolio construction is not an act of nature. It is engineered.
This tension is not theoretical. It sits at the heart of Norway’s ongoing revision of the responsible investment guidelines governing its sovereign wealth fund, one of the world’s largest.
The previous exclusion framework was put on hold after the finance minister, whose ministry sets the fund’s mandate, raised concerns that the Ethics Council was preparing to scrutinise major technology holdings. The minister argued that divestment at that scale would compromise the fund’s index-tracking model and threaten returns. At the same time, the ministry maintains that the mandate remains aligned with the UNGPs and the OECD guidance.
“Even where a relationship is considered crucial, investors are still expected to intensify leverage and seek to prevent and mitigate adverse impacts”
If diversification requires exposure to particular sectors, that reflects a structural decision embedded in mandate architecture. And architecture can evolve.
The uncomfortable truth is that passive exposure to systemic risk remains exposure. If a company’s products reshape digital public space, underpin AI-enabled surveillance, or sit at the frontier of dual-use technologies, then severity – not index weighting – becomes the relevant lens.
The leverage obligation
Even where divestment is genuinely constrained, responsibility does not shrink – it expands.
Under the UNGP framework, investors facing human rights risk are expected to conduct human rights due diligence and publicly articulate clear expectations. Where the scale of harm warrants it, they should co-ordinate collective engagement, vote against management when necessary and support or file shareholder resolutions.
Where these measures prove insufficient, investors are expected to escalate through regulatory or legal channels and – where harms are severe and persistent – to reassess whether continued investment is compatible with their mandate.
The logic is progressive. It is not “engage or divest”. It is “identify, prevent, mitigate, remedy and escalate”. It is certainly not “engage indefinitely regardless of outcome”.
If leverage demonstrably fails, the question cannot simply be whether the company is too large to exit. It must also be whether the mandate itself remains compatible with international standards.
The credibility test
If the largest and most systemically consequential companies are treated as structurally untouchable, the responsibility to respect human rights becomes weakest precisely where risks are greatest.
That would invert the UNGP prioritisation principle, which requires attention to severity of impact, not portfolio convenience.
In effect, the “crucial relationship” argument risks becoming a market-sized loophole: a doctrinal nuance repurposed as a blanket defence for permanent exposure to severe harm.
That cannot have been the intention of the OECD guidance and flatly contradicts the UNGP logic. Nor is it sustainable for a sector that has built its legitimacy on the language of stewardship and commitment to responsible investment in line with international standards.
A more defensible position
A more credible framing for index managers would recognise that, while divestment may be constrained by mandate design, this increases rather than decreases the obligation to exercise maximum leverage. If that leverage fails and severe impacts persist, the mandate architecture itself must be reassessed.
Systemic importance does not dilute responsibility. It magnifies it.
Anthropic’s boundary-setting illustrates that even frontier technology firms recognise limits. Investors and fund managers should not be the only actors insisting they have none.
The real question is not whether certain companies are too “crucial” to divest. It is whether responsible investment remains credible if the most powerful firms in global markets are treated as structurally exempt from meaningful consequence.
If responsibility scales with impact, accountability must scale with size, and with power.
Gerald Pachoud is managing partner of Pluto Advisory and former special adviser to the UN Special Representative on Business and Human Rights.
John Grova is senior adviser, ESG advisory at SpareBank 1 Sør-Norge Forretningspartner. He was previously an adviser to the UN Working Group on Business and Human Rights.
