AI Answers Alone, Without Context, Are Meaningless

When I was younger, my family and I used to refer to my grandad as being a ‘Victor Meldrew’, a character who epitomised the archetypal grumpy old man. Clearly, as I have aged, I have inherited some of those characteristics. I find myself increasingly cynical, wondering to what extent these ‘AI goes rogue stories’ are truly unassisted and aren’t solely designed as a marketing ploy. In my experience using AI daily as my coding copilot, I’m yet to experience an instance where AI loaded with context tried to go out of its way to solve my codebase; if anything, quite the opposite. AI seems perfectly content let alone capable of solving a request I give it but remains oblivious to the codebase around it. I get that’s by design, but if it’s so ‘dangerous’ as the media would have you believe, would it just be sat in the middle of the code waiting for a moment to strike?

If it is, I’ll be in for a nasty surprise and someone prepared to say, “I told you so”. As things stand, AI has increasingly felt more like a tool than a replacement - the blacksmiths hammer as opposed to the blacksmith. Given the media’s habit of convulsing each time an AI does something seemingly of its ‘own accord’, I sit there thinking “I don’t believe it”. Is AI capable in the hands of someone giving directions of attacking things, absolutely; but for now, I feel less worried that Skynet is coming for us all and instead feel somewhat more optimistic about the human factor. For many caught in the mass exodus seemingly centred around big tech firms, stay strong, I believe that you’ll be back to the grind soon enough (I have my fingers crossed for you all).

I wonder whether the layoff is a wave in a technical revolution? In the past something has come along, there’s been a fracas and it eventually subsides as we adapt and evolve. When the typewriter first became publicly available, did everyone go around claiming that ‘pens and paper are a thing of the past’? Are paper and pens gone? No. Granted AI feels like less of a ripple and who knows how far from the epicentre we are. It could still become a full-blown tsunami. The thing is, we’re starting to see the emergence of a trend where companies who have been betting big on AI over a human starting to admit it’s not quite working out as they envisioned. Maybe those making these executive decisions they lacked a little extra context.

I wish I had the answer for what comes next. Unlike an answer from AI, I’m capable of saying “I don’t know”.

“But AI is unlocking all this potential”. It’s true AI is helpful and for those who know their fundamentals, there is value to be had from efficiency gains - though it’s not a silver bullet. I’m actively using AI to help me in my workplace developing a CLI tool which retrieves and analyses email marketing campaigns, a primary goal of it is to identify activity that is likely to be non-human and prevent results from being skewed. Some of the issues include security software pre-scanning emails or checking every link before the recipient ever sees them.

The tool evaluates anonymised behaviours, applies weighted scores and places addresses or domains into confidence bands. There’s a bit more too it, but to spare chapter and verse, let’s say that’s it. To the business I work for, the accuracy of the results matters because not only does it help assess campaign performance for the business, but for clients who have paid us to run campaigns on their behalf. “The output looked plausible” is not really good enough.

As the project has progressed AI has been helpful in testing and implementing known conditions. When prompting the AI with something narrow and specific the result has proven to be excellent, truly excellent. A little back-and-forth, but certainly demonstrating it’s understanding of a problem or objective.

For example:

How can I improve the performance of data retrieval from the mailing engine’s API? Reference the documentation here: […], refer to @folder/file(s) - provide a list of appropriate suggestions, order them by appropriateness and justify which one(s) you would implement and why. Where helpful, provide a SWOT analysis and let’s work through this together. Ask me questions to check/verify any questions or concerns. Don’t make assumptions, refer to the codebase present or ask me.”

The response suggested using a persistent connection rather than repeatedly opening new HTTP requests. A simple fix, but the change roughly halved the time needed to retrieve the results.

A clear question. Useful answer. Measurable improvement.

Broaden the question or leave out the context and things seem to become less impressive.

“How to can I identify bot traffic in the email results”. It can provide a perfectly respectable textbook response: unusual click speeds, identification around IP addresses, repeated activity and improbable patterns, but it’s an answer for answers sake.

Image source, Magnific

You’d think while it ‘thought’ and probed the existing codebase while searching for answers that you would get a response that unlocks a hidden gem, unfortunately, it’s increasingly common to experience nothing revolutionary or even sometimes relevant in the responses. Even when you get a useful answer, the implementation often misses something off. Contrary to the AI hacking narratives, there were no attempts to automatically experiment with things beyond the scope. I didn’t experience attempts to review the detection logic to determine its effectiveness and suggest improvements, identify vulnerabilities, question the scoring model or consider what should happen after the results have been fetched. Before the tin foil brigade or AI evangelists jump in, yes, I’m fully aware that a commercial copy of AI is less ‘savage’ than the unshackled hyper-performance copies the AI providers have at their disposal, but I haven’t had even a whiff of exceptional proactiveness.

I would have at least thought the AI might try expanding some results based on the access to the codebase or even due to ambiguity. Responses in my case could have at times questioned simple things like: Would a user want to check a campaign more than once? Should fetched results be saved and how? Should previously checked results be purged at any point? Should domain confidence be calculated across several campaigns rather than treating every send in isolation?

None of those questions appeared until I raised them.

You see, AI answered my requests, sometimes better than other occasions. With context the results consistently yielded better results, but when vague, perhaps because I didn’t know what to ask or asked to vaguely and I’d end up with a loosey goosey reply. No dramatic hallucinations. No rogue ‘break the box’ code. It simply solved each task in isolation without questioning how it fitted alongside everything else. Oh, and I forgot to mention in both instances I had skills defined, memory files, instructive markdown files to extend known details and provide a richer way to allow an AI to think in depth.

AI can provide answers remarkably quickly. It can write code, suggest improvements and remove a great deal of repetitive work. But it does not always do the thinking behind the task, and you really need to know what to do with the results, else it’s easy to get lost in the rough. After all AI answers the questions but I still had to understand the codebase and the problem.