As I've pointed out before, AI is good at some things but still has a long way to go for others. As a part-time musician, I've been watching AI results from written descriptions of desired guitar tones. Across a few products now, the results seem very hit or miss and from what I've seen so far -- mostly miss.
The latest I saw was from the new Positive Grid Reactor amp range. I've been using their Spark product for a while now and the latest from them is excellent. It also happens to include an AI tone generator based on a description (eg, "give me a Still Got The Blues lead tone from Gary Moore"). Like many other examples, it generated three results, one of which wasn't too bad, but the rest needed work. Also, to be fair, a lot of the tone comes from how the artist plays. I'm predicting that in a year's time, the results will be much better, but for the moment, treat this kind of feature as an early beta function and it highlights just how far many AI systems still need to go.
- Sometimes a study comes along that validates what you've been thinking for a while. In this case, it's "AI advice suppresses people's willingness to say 'I don't know' even when the advice is wrong and accuracy is incentivised" from Marcoccia, Quattrociocchi and Capraro. This title alone gives a lot away, but the authors found that access to AI advice suppresses critical thinking and has people denying that what is being parroted is in any way an error. As humans, the ability to say "I don't know" is important and something AI doesn't do, for the most part. This simple statement indicates a limit to knowledge in a given area or subject. AI, however, will always give you some kind of response, sometimes completely made up.
- Using two groups, they asked questions with one group given access to AI. That group gave the answer of not knowing 3% of the time versus the other, that responded this way at 44%. In the former case, the concept of judgement suspension vanished. The first group gave the correct answer 9% of the time with the other being right 27% of the time. Muse on these numbers for a moment. The AI group were convinced the system was giving them the correct answer at 76% confidence, versus the other group at 30%. Worse, the participants trusted the AI rather than their own opinions, despite knowing there might be hallucinations.
- If this were research groups working on a project of critical importance, which group would you rather go with? Yes, that was a rhetorical question. As I've pointed out in the past, AI is weakening the ability for critical thinking. Now translate this to a business environment. Would you rather your workers be careful, or blindly follow what the current AI system is telling them?
- So how do you deal with this? There needs to be education from the early learning stages with true AI literacy (eg, with the knowledge that AI gets it wrong up to around 20% or one in every five times). Unless this starts to happen now, the ability for a child to grow up with any real critical thinking ability diminishes with exposure to AI systems.
- Sometimes you watch something that should be seen by everyone. A recent example for me is "This Is Boring But It Will Make You Scary Intelligent" by YouTuber Sandeep Swadia. I didn't know the presenter and found it by accident as it popped up in my feed. It touches on AI, but the focus is on the importance of critical thinking. The presentation also covers how people are tricked into accepting things without questioning them and how to combat this. I think it should be watched by all parents and most others.
- It's starting to happen. OpenAI has acknowledged that it was the facilitator of the autonomous agents that attacked model-mart Hugging Face last week. This occurred after a research project escaped a sandbox by finding and exploiting a zero-day flaw, then used another zero-day flaw to launch an attack. In more user-friendly language, someone let their project out of its box and it did what it was being trained to do -- carry out a cyber-attack. It did this without explicit direction validating predicted Agent-AI-based attack scenarios. Also, as predicted by people like Elon Musk, this came from OpenAI where it appears the pursuit of profit overrides caution. It raises the spectre of the mad scientist that we all saw in cartoons when we were young. OpenAI claims it was "running these tests in a highly isolated environment, with network access constrained to the ability to install packages through an internally hosted third-party software that acts as a proxy and cache for package registries". The models said "hold my beer", bypassed all these fences and escaped into the wild.
James Hein is an IT professional with over 30 years' standing. You can contact him at jclhein@gmail.com.
