AI safety

Humans have made many high-risk discoveries such as biological, chemical, nuclear and conventional weapons. Some are in common usage such as high buildings, cars cigarettes, alcohol and ultra processed foods UPF despite their risks. Humanity has developed ways of reducing the danger from these technologies by adding safety features. We have struggled to keep up with dangers of technologies such as social media, smart phones and AI. Social media companies (like tobacco companies) have insisted that they are not responsible for the harm saying they are ‘only a platform’. 

AI companies have suffered two significant challenges, the first was the use of copyright data without permission or payment. This caused harm to many creatives as their style was copied to create new and sometimes inappropriate content without being offered any control. The damage to the internet in reduced quality of material from posting of AI slop was largely due to the lack of an identifier so that copyright and AI generated material cannot be distinguished. 

The second is AI psychosis which is a poorly characterised range of conditions where there is mental health problems related to interactions with AI. I have argued that the pathology is coming from outside the person (from the AI) so it is fundamentally different from normal psychosis where the pathology is driven by brain circuits. If this is confirmed then it will offer a way of differentiating between a person with psychosis who uses AI and a person who becomes unwell due to the AI.

Core Focus Areas for AI Safety

Urgent regulation is required to ensure that these problems are reported, research on how to mitigate is requested and safety systems are put in place. Clear legal responsibility needs to be established for harms. AI safety should focus on the following issues:

Dual use AI

There are increasing numbers of areas where AI has or will soon become superintelligent and can potentially cause harm. Coding abilities has already reached the level where LLMs can easily bypass current security and accelerate hacking. Biological and chemical capabilities are likely to allow bad actors to create known or previously unknown toxins for terrorism. Agentic AI can be used to create attacks on internet sites, phishing attacks or spam to influence people. 

The only effective strategy is to have a warning flag that identifies any inappropriate use of AI. The prompts can then be reviewed to confirm that the actor is malicious and the information passed to the intelligence and security services. As bad actors have the alternative to download an open-source LLM trying to block their actions will just drive them underground. The information gathered from flagging inappropriate prompts can be used to create advice on safe use of AI.

What does safe use of AI look like?

Car safety has developed from bumpers to seatbelts, airbags and sophisticated design features such as crumple zones and automatic braking to protect the occupants and other road users. Many developments were in road safety where road layouts have been responsible for substantial improvements such as on motorways, barriers and roundabouts. The driving test has developed with knowledge, skills and hazard perception demands increased. Driving awareness and speed awareness courses have replaced simple points of the licence where appropriate. 

Adapting Cross-Industry Fail-Safes

The AI Kill Switch is concept that became popular in political circles in 2026 and has initial attraction. The idea that a datacentre could be disconnected if the model behaved in a dangerous way is both simplistic and extreme. However if the threat was sufficient then this local disruption might be better than closing the internet. The ‘AI Kill Switch’ represents the types of fail-safe systems that are already used in nuclear energy. 

Circuit Breakers are a similar idea from financial markets where the AI is automatically frozen if its behaviour becomes abnormal. The limitation on this type of safety process is that it is unclear whether it will be possible to detect abnormal behaviour. It is however worth putting in place circuit breakers that pause the system so that humans can check it even if they are manual. 

Structural engineers use a Factor of Safety (FoS) which is the multiple of maximum stress predicted. A similar concept can be used in medical systems where they can be tested in noisy environments and the falls in accuracy determined. If a MRI reporting software gets confused with artefacts then the humans can disregard the AI report in a poor-quality scan. 

Early warnings are likely to precede any catastrophic breakdown which has led to the concept of a near miss from aviation. A ‘Just Culture’ protects the person reporting a near miss or minor technical problems. We will need AI "Near-Miss" Registries to allow people to report when and AI behave erratically or attempts a prohibited action. The black box recording is a ledger that could be used to understand what happened when an AI goes rogue.

Conclusions

If universities do not take responsibility for AI safety training then the government should ask AI companies to help build a course. The examination should run and marked by UK based experts and scored using grades 1-9 like GCSEs. Ofsted should inspect the quality and encourage gradual improvement to include new threats and developments as they occur. 

Governments have had mixed success in protecting their populations from the harms of technology. Car safety has been an extraordinary success but alcohol risks are largely an ongoing failure. Despite there being simple and straightforward steps that could make AI safer there has been little progress. Steps such as watermarking, bias testing, red teaming and human in the loop have been ineffective. AI companies should be made legally responsible for any harm their products cause unless they take AI safety seriously.

August 2026