The eve of the US and Israel’s attacks on Iran was the deadline for artificial intelligence developer Anthropic to capitulate to the US Department of War’s terms for unfettered access to its AI tools or risk losing a $200m contract. Anthropic held fast, releasing a statement that said, without guardrails against use for mass domestic surveillance or autonomous weapons, the firm would not agree to the DoW’s requests. In response, Anthropic has been designated a ‘supply chain risk’ by the DoW.
This stance aligns with red lines drawn by Dario Amodei, chief executive officer of Anthropic. In a blog post, he outlined the risks posed by the proliferation of powerful AI and its subornment by state actors. While Amodei’s policy responses run the gamut from actionable and necessary insights to idealised spin-offs of his own belief system, he is crystal clear on the potential consequences of AI abuse by governments.
Anthropic’s DoW stance highlights the realities of AI’s increasing importance in governments’ toolkits and the associated risks. Uncritical and unrestrained use of the technology, without adequate guardrails against misalignment, abuse and drastic economic influence will have severe consequences. Radical and comprehensive political action will be required to avert these but, at present, there is little alignment between policy-makers and AI developers on what form this should take.
Preventing abuse of AI
Amodei points out that powerful AI is susceptible to misuse, empowering terrorists to commit much worse atrocities by giving them access to information on bioweapons or cyber-attacks, or by empowering governments to engage in mass surveillance. AI transcription capabilities and sorting power makes it possible to centralise records of conversations and actions of individuals, creating the capacity for flagrant abuses of human rights.
A civil-liberties approach to AI regulation in this case prioritises the privacy and autonomy of individuals and presents a framework through which legislative and policy actions can guard against this kind of data harvesting.
There are two dimensions to the prevention of government abuse of AI: international and domestic.
For the curtailment of abuse of AI by foreign powers, Amodei recommends a ‘crimes against humanity’ framework for use of AI technology for autocratic means. This would be supplemented by a global ‘taboo’ on these uses, coupled with a hawkish approach to trade, restricting the sale of cutting-edge AI inputs to potentially autocratic regimes. This requires a degree of international co-operation but, since the number of countries producing such equipment is small, it may not be unreasonable.
Domestically, the situation is perhaps more challenging. When the government is the customer, legislation is insufficient. While Anthropic may have made a principled stand on this issue, the rapidity with which a competitor took on the DoW contract demonstrates the difficulty of setting ethical standards unilaterally.
Misalignment over goals
The threats Amodei identifies are not limited to the misuse of AI by states. He also highlights the risks of ‘misalignment’ of goals between the AI and its user. While Amodei rejects the idea that hysteria-tinged predictions of rogue AI models are inevitable, he takes the concerns seriously and even acknowledges that legislation will be necessary to steer AI development to a safe course.
As well as calling on labs to make alignment a key part of their training approach and highlighting the need for research on identifying misalignment risk early, Amodei encourages governments to take an active approach: live model monitoring for both internal and external use with legally mandated disclosure of misaligned behaviours.
Since admitting bad behaviour by AI is not in developers’ interest, Amodei is quite right to call on lawmakers to develop proper supervisory capacity.
At present though, neither the other leading AI companies nor the US government appear to be aligned with Amodei’s approach.
Mitigating economic impact
Among the most important areas of AI policy is the question of how governments should respond to the economic disruption it might cause. Amodei lays out eight recommendations, with voluntary wealth redistribution as a ‘culture of philanthropy’ appearing twice. These fall into three categories: corporate responsibility, individual responsibility and political responsibility.
Monitoring the impact of AI on job displacement is certainly a vital step for assessing the economic impact of AI, but when it comes to mitigating it, Amodei’s recommendations run into problems. He encourages companies to avoid laying off staff and continue to pay employees once their productivity has been superseded by AI. Without a legal incentive to do so, this recommendation clashes with the reality of corporate incentives.
In a similar vein, Amodei warns against the potential for rapid wealth accumulation through large language models and AI’s impact on the job market. He cites Anthropic’s own policy to donate 80% of its wealth and company shares towards charitable causes. However, tech-driven revolutions are not new and each has increased the rich’s share of global wealth. Wealth accumulation due to AI cannot be rectified by the voluntary actions of the richest few. Since 1975, inequality in the US, measured by the Gini coefficient, has been steadily increasing, hovering around 41.8 in 2023, from 35.6 in 1975. In order for voluntary philanthropy to work this time, there has to be radical departure from the economic impact of technological innovations from the last 50 years.
Amodei also recommends a macroeconomic shift. While not recommending a specific tax redesign, ideas of stronger progressive taxation or a tax levied specifically against AI companies point to the necessity for comprehensive macroeconomic policy that takes job loss into account. Facing a threat of severe economic disruption – namely mass job loss affecting entry-level workers and the white-collar middle class – there must be a macroeconomic approach on a national scale.
Domestic legislation will not be sufficient. Amodei recognises the necessity of state-intervention away from concentrating economic power in the hands of AI producers, but this requires a global effort as well. Concentration of wealth via AI to a handful of people within one country will reflect domestic worries of lack of competition and innovation on a global scale.
This requires policy-makers and public opinion to take seriously AI development and to critically engage with what this technology can, and should, do. Doomerism is not necessary, but comprehensive political action is.
Jordan Nann is Accounts and Content Executive at OMFIF.
