Original document: https://ari.us/wp-content/uploads/2025/08/Agentic-AI-and-Liability-The-Stick-the-Carrot-and-the-Net.pdf
AI “agents” are pushing AI into the real world, and this report explains how law and policy can deal with the harms they cause using three tools: a stick (liability), a carrot (rewards for good governance), and a net (no‑fault compensation).
Background: What are AI agents?
Most people know AI as chatbots or prediction tools that answer questions or complete text, but “agentic” AI can act on instructions, take multiple steps, and interact with other systems without constant human control. These agents can book tickets, reply to customers, write and run code, or manage workflows, which means their mistakes can look less like bad advice and more like real‑world actions that cause financial, physical, or legal harm.
Why existing rules are not enough
Traditional software law treated programs as tools that humans directly operate, so liability often stopped with the human user or was limited by contract terms and platform immunity. With autonomous AI agents, responsibility becomes murkier: a bad outcome might involve the model provider, the company that integrated it, the business deploying it, and the end user, making it hard for victims to prove who was at fault and discouraging firms from investing seriously in safety.
The stick: stronger liability
The stick means making it easier to hold companies legally responsible when their AI agents cause harm, so they have strong incentives to prevent those harms. The report suggests ideas like shifting the burden of proof toward developers when they have ignored clear safety practices, and requiring more transparency about model design, training, and testing so courts can decide what “reasonable care” should look like.ari
The carrot: rewards for good practice
The carrot is limited legal protection (safe harbors) for companies that follow robust governance standards, such as independent audits and detailed documentation of risks and controls. Examples include proposals where certified systems or tools used by licensed professionals receive some liability shield, but only if they meet strict transparency and safety requirements.
Risks of overusing the carrot
If safe harbors are too broad or too easy to obtain, audits can become box‑ticking exercises and “rubber stamps,” leaving risky systems effectively immunized while the public bears the cost of failures. Because AI safety science is still immature and standards are in flux, the report warns that large immunity schemes could lock in weak practices and undermine both deterrence and accountability.
Calibrating the carrot
To avoid these pitfalls, the report recommends narrower protections, such as capping damages instead of eliminating them, limiting safe harbors to specific low‑ or medium‑risk uses, or shielding developers mainly from claims by informed users but not from harms to third parties who never chose the system. These calibrated approaches aim to reward proactive safety without erasing meaningful remedies for people harmed by AI agents.
The net: no‑fault compensation funds
The net is a safety‑net model: victims of AI‑related harms receive compensation from a dedicated fund without needing to prove anyone was negligent, similar to vaccine injury programs or some oil spill funds. This can be useful when it is technically hard to untangle who exactly caused the harm or when traditional litigation would be too slow and costly for individuals.
Challenges in building the net
Because AI deployment is new and diverse, there is little solid data to price risk, making it hard to set contribution rules for companies or to keep the fund solvent over time. There is also a danger that firms rely on the fund instead of investing in prevention, so the report suggests starting with narrow pilots, strong incident reporting, and mechanisms to recoup payouts from clearly responsible companies.
Why all three tools are needed
The report argues that no single approach can manage AI agents on its own: liability (the stick) deters reckless behavior, governance‑for‑protection (the carrot) encourages better practices in advance, and compensation schemes (the net) ensure people are not left without help when harms occur despite precautions. Policymakers are urged to design a mixed system that uses all three, tuned to different sectors and risk levels, so society can benefit from powerful AI agents without accepting unchecked and uncompensated risks.
