A Letter From a Consumer

Two ways, mostly. The first is money. Where consumers spend must be where consumer interest lies, and that’s hard to argue with. Spending correlates with demand, demand justifies supply, and the industry feeds itself. Nothing wrong with it. It makes perfect sense.

The second is research. Surveys, panels, focus groups — all producing a consensus. And a consensus is an average: the middle of the range of people who are already interested in the thing being asked about. That aggregate decides where the R&D money goes. Also perfectly reasonable. You want to know there’s an audience before you build.

If people knew what they wanted, they’d go and buy it. They’d search for it religiously. An aggregate can only ever tell you about things people can already imagine.

People are adventurous — all of us, collectively. We’re always looking at what’s just over the ridge. When a space gets crowded, some portion of us gets restless and wants a journey.

That does two things. It tells us what’s out there. And it tells us what’s terrible out there, because the person who goes looking comes back and says no, really, don’t.

We’re also intensely group-minded, and that’s why it works. As humanity we have always benefited from the few who are bold enough to go striking off in a direction, so sure it’s going to be amazing. The rest of us may not be on board right away — but because we are a social network, if it succeeds, we notice. And then more people go, until the whole crowd has moved. It runs identically in reverse.

So this site is one person coming back and reporting.

In 1997, Dragon Systems released Dragon NaturallySpeaking — the first product that let you talk continuously and have it come out as text. It shipped with a headset. Its big marketing pitch was exactly what it sounds like: you talk, and it writes for you.

I was young and there wasn’t money for it, so I never got it. Within a couple of years I stopped hearing about it.

Later, Word gained dictation and I was thrilled — until I used it. You had to say “period”, “new sentence”, and it still misunderstood you constantly, and you’d repeat the same word over and over. I remember thinking I could have finished the whole paper by now. It wasn’t a bad product so much as an unfinished one. What I’d seen advertised in 1997 was continuous speech; what I actually met was the older, word-at-a-time kind.

Years on I was making videos and I needed voices. Someone with a lovely voice was generous with her time, and I felt like a nuisance every time the script changed — which was constantly, because nobody cares about the script until you start recording, and then everybody does. Every change cost weeks of scheduling.

So I got interested in the reverse problem. If there’s speech-to-text, there must be text-to-speech. Or you could record a voice once and treat it like a typewriter.

There was, and it was awful. Robotic. I only wanted it as a placeholder to prove the concept, but nobody could hear past the robot. Later, Amazon launched Polly in November 2016 and the demo sounded extraordinary. I signed up immediately. What came back was robotic again. The smooth voice in the demo was not the voice in the product.

Everyone around me had gone stale on the idea. I hadn’t.

I want to be honest about this part, because it’s the reason I take reviewing seriously now.

For years, “AI” was a buzzword to me. White noise. When ChatGPT appeared I brushed it off without ever using it — I assumed it was Siri as a web app. My mental model of the current state of AI was Amazon Polly: cobbled together, imperfect, oversold.

Then NPR’s Planet Money ran a series in which WellSaid Labs cloned the voice of former host Robert Smith — and put the synthetic version on air against live hosts who played off it like a colleague. The running joke was that it was coming for their jobs. When I heard that voice I couldn’t believe it was synthetic.

There was one tell, and it’s worth explaining because it’s exactly what was still missing. A real speaker’s pitch rises and falls constantly — you lean on a word to stress it, your tone lifts when you’re unsure — and all of that carries meaning before the words do. That voice didn’t. Put a number on it: human speech ranges right across the scale, and that voice stayed between four and five, every now and then went up to six, and never really strayed. It’s still the tell today, though they’ve got much better at it.

But the timbre was him. That was real. And my honest reaction wasn’t “this is amazing.” It was: what else have I been wrong about?

So I started trying everything. Properly this time. Though for a good while after that I was still using ChatGPT as an image generator and nothing else — it wasn’t until a colleague who’d been taking AI classes sat down and explained what it actually did that I understood. I was floored. It’s a capable agent, and it’s helped me enormously since.

That’s a dismissal I got wrong twice, and I’d rather say so than pretend I saw it coming.

Apple shipped Siri in October 2011 and called it an assistant. Everyone else followed: we have an assistant too.

No, you don’t. You have a tool.

What was actually announced was Google, if Google lived in your pocket and could set a timer, add a calendar entry and send a text message — often badly. And the competition was about doing that slightly better. Ours handles messages. Ours talks to the washer and dryer. Ours does the maths problem.

Is a tool useful? Absolutely. If you use your hands all day and a question pops into your head, that’s genuinely valuable. I’m not saying it isn’t.

And that means being connected to your personal data. None of them understood that, and worse, none of them were talking about it. That’s the part that still bothers me.

It isn’t that the assistant would be malicious. It’s the route. To do anything genuinely personal your data has to travel, and it passes through some of the least secure points on the internet. It isn’t the data centre the AI lives in, and it isn’t the pipe to it either — it’s at all the interconnected points, the points where your data gets decrypted.

I’d love a real assistant. I’d trust one with my life admin tomorrow if anyone were honest about that problem. Nobody is.

And current AI has the same blind spot, in a new suit. If you’re lost and trying to get a project done, it just does what you ask. It doesn’t stop and say are you sure this is what you want? It will carry you miles from the idea you started with, executing every step perfectly, and never once help you get where you were actually trying to go. That is the identical failure to the survey: it answers the question you asked, and has no way of knowing about the thing you couldn’t articulate.

I don’t know anybody in big tech. I’m one person. So this site is me shouting anyway. It reminds me of ham radio operators — never many of them, a genuinely niche set of skills, sharpening away, and a fair amount of the time sitting on a channel talking to nobody in particular, hoping something comes back.

That’s what this is. I’m a writer; this is the thing I can do. It never crossed my mind to make a YouTube channel — I commend the people who get all dolled up every day, but you have to have a passion for that, and it’s a whole lot of extra work. It isn’t mine. And I figured there had to be other people who think the way I do.

The other half of the idea was a test. If AI is coming for writing jobs, fine — let it have mine and see if anybody notices. So most of the early site was me telling an AI to write, starting with Rytr and later with Claude. I reviewed the tools I actually ran, which is why the reviews here are about things I genuinely used. The result: no traffic. Almost nobody came, except the scrapers.

But something unexpected came out of it. I love writing, and my problem has always been the same. I get an idea and I can see the shape of it — it’s a wave, and it carries me. Then I stop to check a spelling, or whether I used the wrong there, and the wave begins to leave me. I hurry to catch it, and then everything from school arrives: how many sentences is that, did I do an open and a close. And I look up and the wave is gone, and I have no idea what I was doing.

So I tried talking instead. I talked for a long time, then asked Claude to pull out what mattered.

You know those sculptures where it’s a bunch of garbage suspended from wires, and if you look at it just the right way, you see the Mona Lisa? That’s what it felt like. Reading back the points it pulled, I could feel the wave again, and everything started to line up. It might be a piece of garbage — that isn’t the point. The point is that this is what I wanted to say, and it snaps into place.