Some (possibly stupid) questions about AI

I am very much a beginner when it comes to using AI. Like many, I’m hearing the headlines, but don’t really have much personal experience. I know those in the tech and business world are seeing big impacts. We had some helpful orientation to the issues a few weeks ago with a guest speaker from the Faraday institute (What does AI have to do with Christianity?). But I know that others know far more than me. So if these questions are missing the point or have easy answers- please take pity on me and educate me.

1. Who is liable if I ask AI to do something legitimate, but in trying to achieve that aim AI does something illegal or damaging? With a car, the manufacturer is not liable for the accidents I have when driving, unless I prove that there is a steering fault or something similar. But with AI, I might ask AI to find information or achieve a goal that is legitimate, but the AI might do something illegal or harmful (e.g. hacking another system) to achieve that goal. Who is liable? If it is the company, I imagine that some pretty big lawsuits will come their way. If it is the individual who does not understand how the AI will work to achieve the goals, then I imagine it will make companies and individuals pretty cautious about using AI. Issues of liability seem to have held up self driving cars. Is there a reason the liability issue is not delaying AI rollout?

2. How important will it be to have an AI that really knows you (so you are stuck with one provider), or will AI’s be interchangeable? Does AI have to spend a year tracking everything you do to be able to give answers with the context and level you want? Or can a relatively compact set of instructions get any AI to give useful answers? If the former, AI is something you have to invest in order to function at a high level. If the latter, it is something you can wait till the models improve without missing out too much. And in the former, AI providers getting you hooked are gaining lifelong customers with a deep moat (and hence a high value). But in the latter, AI become a like a utility with price causing frequent switching. Is AI going to lead to Microsoft/Apple/Facebook type gains from being dominant, or a service more like broadband?

3. Can AI overcome the technological inertia affecting areas like health and education? UK healthcare is notorious for how poorly tech has been joined up. Accessing test results, booking systems, and general use of technology has lagged far behind what tech appears to make possible. You might argue the same with education. So a potential positive outcome is that AI might enable leaps over current bottlenecks, and inefficiencies. If AI simply listens and watches a class, and at the end notes children who were struggling and who get special help, without the teacher having to write notes, might more accurate pictures of children’s needs and best responses emerge? If AI enables patients to be guided smoothly through the whole treatment pathway and to give feedback where the treatment is not good, might healthcare improve substantially?

4. Will 90% of AI gains be lost to competitive AI and additional administrative sludge? If institutions and individuals all use AI, is the result that the majority of AI usage is in dealing with other AIs, with the net result being that most Ai is not actually moving things forward, but you are forced to use AI to get through the AI defences opposed to you? In the charity sector, getting skilled at getting grants is a vital skill, but you can end up with as much being spent by the charity and the donors on managing the process of getting the grant as directly on the activity of the charity. Does much of everyday AI (as opposed to top level research AI) more or less cancel out?

5. Will AI lead to a return to closed networks and in person community? People seem to want to spend much of their time in safe spaces with people they trust. If AI amplifies the existing problems with bots and scams online, does the internet become divided into a range of walled gardens, while the open internet becomes a wilderness? If AI makes trusting any online communication, even video calls being fakeable, do we return to in person meetings?

6. If Government mandated GDPR forced basically all organisations into the cloud, and then AI makes the cloud unworkable, can we sue the government? This last one is a bit cheeky. In 2001, the government encouraged the purchase of diesel cars to help reduce CO2 emissions. In 2015, the government began penalising diesel cars for their effect on local air quality. Having encouraged an expensive purchase, the government then destroyed the resale value of diesel cars. The GDPR regulations have basically forced organisations to pick a cloud provider (Microsoft or Google mainly). If AI makes the could untenable, there is going to be a lot of unpicking that work. If the governments laws have made organisations more vulnerable, can we sue the government?

I welcome answers to these questions from people who know more than me.

(For an attempt at more constructive engagement, see What virtues and practices might help us in an AI world? – Jots and Scribbles)

3 thoughts on “Some (possibly stupid) questions about AI”

  1. In brief response:
    1) I don’t think that there is a single answer to this question, but in talking about driving a car – you’ve picked my favourite metaphor. Car manufacturers are highly regulated and are responsible for providing a safe instrument. Drivers are highly regulated and are responsible for the use of that instrument. If you want an example, this is a case study on an individual who poisoned himself with bromine based on advice that he obtained from ChatGPT: https://www.acpjournals.org/doi/10.7326/aimcc.2024.1260.
    2) There’s a couple of reasons why it’s good to invest deeply in a single AI ecosystem. One is data access: the advantage of Copilot (Microsoft’s AI) is that it operates within your existing systems – for example, it can read your emails. The other is building technical capability – if you want to repeat a task, you can create an agent, so that it can keep doing that task, and learn your preferences. Conversely, switching ecosystems is good to avoid blind spots.
    3) That is the vision in the Sudlow review: https://www.hdruk.ac.uk/helping-with-health-data/the-sudlow-review/. There’s a lot to overcome, but that is certainly the intended destination.
    4) I’m not sure about your phrasing, but we are certainly seeing some of this in science. AI enables mediocre papers to be written faster than ever. Good scientists spend more time reviewing and trudging through mediocre work.
    5) I haven’t seen this as yet, but it’s possible. One trend in this direction that I’ve seen is much less trust in giving people access to systems. Taking on a summer intern has always provided limited benefit to the company. If taking on a summer intern means opening up your system to one more user, then it also comes with a large risk. So companies are reluctant to take on interns.
    6) Afraid I’ve not had this experience, so wouldn’t be able to comment. But I suspect the answer is no.

    Liked by 1 person

  2. Large Language models come in two modes today

    1. Chatbot – you ask it questions. It can research but the only action it can take is replying to your questions.
    2. Agent – it’s plugged into your computer/network and can take actions like tidying up your network drive or hacking your friends.

    I have a fair idea of the answer to these for the chatbot, but I’ve never used an agent, so I’ll just bracket some speculations on agents.

    1. You are liable for anything bad caused by you acting on false information provided by a chatbot – that’s the little notice at the bottom telling you to check before you act on it. AI companies are nervous about losing a court case on someone committing suicide following AI advice, and are trying hard to stop this happening, but we’re not very good at controlling our AIs.
      (Responsibility for bad agent actions will need to be tested in court, but will probably resolve to complex mess of responsibility to different parties in different situations.)
    2. It isn’t important to have an AI that really knows you. AIs read context really fast. In a business context you do have to build the set of documentation that the AI can read (or set up the APIs for an agent to act through), but once you’ve built it a different family of AIs will read it at blistering speed. So you do lose out by not starting the journey of getting AI ready, but you don’t lock yourself in to one provider.
      The AI monopoly that people are worried about is recursive self improvement. One of the things AI is best at is writing code. We’re now at the point where AIs are borderline better than humans at developing the next family of AIs. One AI company that got a bit ahead and used their better AI to build an even better AI which they then used to build an even better AI etc. could end up with an AI so much better that we all wanted to use that one.
      At the moment the leaders are Anthropic (Claude and friends. The best at coding, at least until recently, oddly bad at drawing pictures), and OpenAI (ChatGPT. The best at frontier maths problems.)
    3. No. AI is even more dependent than humans, statistical models and ML on having access to lots of data. The barriers to adoption that are holding back all the other potential tech offers will also hold back AI. A lot of the talk in the business world at the moment is about the hard work of building the data infrastructure to enable you to take advantage of all those sexy AI capabilities. AI might make the benefits big enough that you actually overcome the inertia, and incidentally pick up all the other benefits we’ve been talking about for years.
      (You might be able to overcome this by supercharging your IT department with AI coding agents.)
    4. Totally speculative, but we probably ought to assume that near 100% of the gains in adversarial settings (who gets the charity grant) are absorbed by the competition, and near 0% in non-adversarial settings. Of course if all charities became more effective then people might be more willing to give to charity.
      The one that’s been most discussed is hackers vs. security teams. The current thinking is that in the long term AI benefits the security team (before releasing this software I ask AI to check it for vulnerabilities, and fix all of them. My software is now incredibly hard to hack), but for the next few decades it benefits the hackers, as there’s lots of software out there that was released before AI, or before AI got good at finding all the vulnerabilities.
    5. No one knows, but this sounds like wishful thinking to me.
    6. How’s your case against the government for making you buy a diesel car going :-).
      (More seriously – if you’re connected to the internet but not sophisticated enough to implement GDPR then the cloud providers are more secure from AI cyber attacks than you are. Businesses that are sophisticated enough to manage GDPR are cooling on the cloud, mainly because it’s so expensive.)

    Liked by 1 person

Leave a Reply