Why Imperfect Actors Still Move Technology Forward
An escalation ladder and intent-impact matrix for judging AI companies and countries - not by purity of past actions, but by the direction in which their principles evolve.
Authors
In tech policy discussions, people often focus on past events. When a company does something positive, critics quickly bring up old mistakes or accuse it of acting out of self-interest. This is understandable since tech firms have lost trust over time. Still, demanding perfection all the time can actually slow down real progress.
A recent example involves Anthropic and the U.S. Department of Defense. Anthropic requested rules for how its models would be used in defence and accused Chinese models of copying its systems. Critics soon highlighted Anthropic’s past inconsistencies, shifting the discussion from the value of safeguards to whether the company could be trusted.
There is a problem with this way of thinking. It treats companies like people, assuming that past mistakes mean they cannot work for better standards now. But companies are not individuals. They change, adapt, and respond to incentives, rules, and competition. In fast-moving fields like artificial intelligence, new norms often come from this messy process. If we demanded a perfect record before allowing new rules, progress would stall.
The real challenge is not to decide if someone has always acted perfectly. Instead, we should ask if their actions are helping things move in a better direction for everyone. What matters most is making progress.
To understand this progress, picture a ‘Principles Progression Ladder.’ This ladder shows that companies rarely go from acting only in their own interest to being fully responsible all at once. Instead, they move up step by step, responding to incentives, criticism, and real-world pressures.
Principles Progression Ladder
| Rung | Stage | Description | Example |
|---|---|---|---|
| 1 | Opportunistic Advantage | Companies invoke principles only when they directly serve competitive interests. Ethical language appears mainly as strategic signalling. | When multiple AI firms criticised open model releases by competitors while aggressively expanding their own proprietary platforms. |
| 2 | Defensive Adjustment | Firms adopt limited safeguards primarily to manage reputational risk or avoid regulatory backlash. | After early misuse concerns, OpenAI delayed the full release of GPT-2 in 2019 citing potential misuse risks. |
| 3 | Strategic Compliance | Safety and governance mechanisms begin to be operationalised inside organisations, though still motivated by mixed incentives. | Google establishing internal AI ethics review processes after controversies around facial recognition and bias. |
| 4 | Competitive Safety Signalling | Firms emphasise safety frameworks partly to differentiate themselves from rivals and build trust. | Anthropic highlighting “constitutional AI” and structured safety evaluation frameworks. |
| 5 | Normative Advocacy | Companies push for industry-wide testing regimes and safety rules across competitors. | Major AI labs supporting global safety discussions around the AI Safety Summit. |
| 6 | Systemic Stewardship | Firms actively support shared governance structures and safety infrastructure across the industry. | Cross-industry commitments for frontier AI testing cooperation with governments and international safety research networks. |
The ladder works because incentives shift over time. At first, companies often talk about principles mainly to help themselves. Even this can be useful, since it brings safety language into the conversation. Governments can encourage these talks without strict rules. For example, countries can promote transparency statements, voluntary reporting on AI risks, or sharing training practices. These softer incentives make companies compete on safety, not just technology.
The second rung comes when reputational or regulatory pressure starts to shape behaviour. Companies begin making defensive changes to avoid crises. Governments can support this stage with targeted oversight. Regulatory sandboxes, voluntary audit frameworks, and early incident reporting can help companies consider risks without adding heavy compliance burdens right away. When OpenAI delayed GPT-2’s full release, governments could have supported this by setting clear expectations for staged deployment of powerful systems.
At the third step, safety processes become part of daily operations. Companies set up internal review teams, governance boards, and testing systems. Countries can speed up this shift by creating certification programs or audit standards for advanced AI models. If companies know that following safety rules helps them win contracts or get approvals, they have more reason to make these practices standard.
The fourth step is when companies use safety to show they are reliable. They start promoting their governance systems as proof. Governments can help by publishing safety benchmarks or model evaluation standards. Public rankings or scorecards might seem bureaucratic, but they can make safety something companies want to be recognised for.
The fifth rung marks the shift from internal governance to ecosystem-wide norms. Companies start advocating industry standards and cross-company testing frameworks. This stage is where governments should begin coordinating internationally. When companies publicly support shared testing regimes or transparency commitments, countries can formalise these into voluntary international agreements or regulatory baselines.
The final step is about caring for the whole system. Companies see that a stable tech environment benefits everyone, including themselves. Governments can help by creating international groups or shared safety systems. Global evaluation centers, joint research, and worldwide AI incident reporting are all helpful at this stage.
Looking at the progression ladder, Anthropic’s recent actions seem to fall between the fourth and fifth steps. By insisting on safeguards for defence use of its models, Anthropic showed its focus on safety, both in its internal governance and as a public stance in the debate over military AI. By setting clear conditions for deployment, the company also pushed the conversation toward new norms. Anthropic was not just managing its own risk; it was trying to shape expectations for how advanced models should be used in national security. Its criticism of alleged distillation practices by competitors can be seen as commercial defence, but it also raises questions about training integrity and accountability. In both cases, these actions aim to influence the broader rule-making environment, not just internal practices.
While the ladder shows how things change over time, policymakers also need a way to judge actions as they happen. The Intent-Impact Matrix gives another way to look at these decisions.
Intent-Impact Matrix
The logic of the Intent-Impact Matrix rests on separating two questions that are often wrongly collapsed into one. The first axis examines intent - whether an action is primarily driven by strategic self-interest or by a stated commitment to broader public welfare. The second axis examines impact - whether the consequences of the action actually advance the collective interests of society or generate negative externalities. These two axes are deliberately orthogonal. A company may act out of pure self-interest and still produce outcomes that benefit the ecosystem, just as an actor may claim noble intentions while producing harmful consequences. By keeping intent and impact independent rather than intersecting dimensions of the same moral judgement, the matrix forces analysts to avoid simplistic narratives. The framework therefore answers a practical policy question: how should governments and observers react to an action, regardless of the motives behind it? By disentangling motives from outcomes, policymakers can reward beneficial behaviour even when incentives are mixed, while still intervening when well-intentioned actions produce systemic risk.
| Positive Impact on Humanity | Negative Impact on Humanity | |
|---|---|---|
| Strategic Intent | Constructive Pragmatism | Strategic Containment |
| Public-Interest Intent | Principled Leadership | Misguided Idealism |
“Constructive pragmatism” describes actions where companies act for their own reasons, but the results still help society. For example, when leading AI companies agreed to outside safety testing under government pressure, they wanted to improve their image, but the outcome was positive. It raised standards for risk checks across the industry. Governments should support this with contract advantages, tax breaks for safety research, or public awards. The goal is to encourage good results, not judge motives.
“Principled leadership” is when both the intent and the outcome benefit the public. Joint AI safety research projects are a good example. When companies fund independent research on model safety or standards, they help share knowledge across the field. Governments can support this by co-funding research, building public-private partnerships, or offering international grants for safety work.
“Strategic containment” is when companies suggest rules or safety ideas that help themselves but make things harder for competitors. For example, big tech firms might push for rules that only large companies can meet. Governments need to be careful in these cases. Instead of rejecting these ideas outright, they can adjust the rules to keep safety while making sure smaller companies are not unfairly blocked. One way is to set different requirements for companies of different sizes.
The last area, “Misguided Idealism”, is about actions driven by openness or tech optimism that end up creating big risks. For example, releasing powerful open-source models without enough safeguards can lead to scams, misinformation, or cyberattacks. Governments should not stop openness completely, but should encourage responsible open models. They can fund secure open model systems, support safety tools, and require risk reports to reduce risks while still allowing innovation.
When evaluated through the Intent-Impact Matrix, Anthropic’s actions most plausibly fall within the quadrant of constructive pragmatism. The motivations behind requesting safeguards in defence applications are unlikely to be purely altruistic; they also protect the company’s brand positioning as a safety-focused AI developer and reduce potential reputational risks associated with military misuse. Yet the impact of raising such safeguards in public discourse is largely positive. It forces governments, defence agencies, and other AI developers to confront the governance challenges surrounding military deployment of advanced models. Even the dispute over model distillation reflects this dynamic. While it clearly serves Anthropic’s commercial interests, it simultaneously raises questions about the integrity of training practices in an ecosystem where models increasingly learn from one another. In other words, the motivations may be strategic, but the effect is to introduce useful norms into the debate. From a policy perspective, actions that fall into constructive pragmatism should not be dismissed merely because incentives are mixed. Instead, governments should recognise the beneficial systemic effects and reinforce them through clearer standards and oversight mechanisms.
Together, the ladder and the matrix give two ways to judge how AI companies and governments act. The ladder shows long-term changes and how norms shift over time. The matrix focuses on what each action does, separating the reasons behind it from the results.
History shows that most tech governance systems develop slowly. For example, aviation safety standards took decades of accidents, investigations, and cooperation between companies and regulators to develop. Financial rules also grew through cycles of crisis and reform.
Artificial intelligence is now at a similar stage. Governments, companies, and researchers are figuring out what risks are acceptable as things happen. In this situation, demanding perfect behaviour from anyone does not help.
A better approach is to stay practical but alert. Policymakers should reward progress up the principles ladder and judge each action with the intent-impact matrix. They should encourage constructive pragmatism, support principled leadership, handle strategic containment carefully, and address misguided idealism with smart rules.
For countries planning AI strategies, this approach has clear benefits. Instead of judging companies as if in a courtroom, governments can treat tech governance as a system of incentives. By shaping these incentives carefully, they can gradually guide companies and entire industries toward safer and more responsible innovation.