What ‘AI Mark’ taught me about the realities of the new technology
4 min read
In a future that no one asked for, but many predicted, I took the plunge and turned myself into the first AI MP.
It was the world’s first robo-MP, or just another useless robotic politician, depending on who you ask. The results were both predictable and, I’d argue, useful.
So, why did I do this? Jeremy Smith is the head of Neural Voice, an AI startup based in my constituency of Leeds South West and Morley. He suggested my constituents would benefit from having access to two versions of Mark Sewards: the living one and one that never sleeps. Given that my children are one and three years old, I’m not sure that’s the most useful distinction.
The plan was simple enough. AI Mark would emulate my voice and answer casework and policy questions while my office was closed. It wouldn’t replace the work my team or I did, but it could add value as a clever voicemail, and I’ll go to almost any length to support a local business. So, I said yes.
AI Mark started with ‘GPT-4o mini’ as his base, which we built on by feeding him my stances, speeches and typical solutions to casework questions. A text-to-speech component converts both questions into text and answers into speech, with a voice synthesiser to tie it all together.
I spent about two hours recording my voice with Jeremy. He asked a range of questions to capture emotional responses and the full vocal range. Listening to AI Mark now, I was too posh and direct at times; some people found him a bit angry when asked questions he couldn’t answer.
Before launching AI Mark, we tested him extensively. It doesn’t take long to find examples of AI models that have been manipulated into saying horrific things, and I was keen (nervous even!) to ensure that that didn’t happen here. After finding a starting point I was satisfied with, we launched him in August.
I knew journalists would try to break it, so we sent AI Mark to them first. He held up reasonably well, although they were disappointed he wouldn’t give opinions on opposition politicians or my own Prime Minister. PoliticsHome did get him to produce a haiku about Nigel Farage with the wrong number of syllables. I quickly fixed that.
The reception on social media drove most of our traffic. Some people spent hours trying to break him. One person from Ukraine had long conversations with him, trying to get him to declare support for one repressive regime or another. But AI Mark held firm in his love for democracy.
This is where the limitations of the technology became clear. To have useful conversations with AI Mark, the limits on what he could say – the guardrails, as I call them – had to be lower. But lowering them even slightly increased the risk that AI Mark would produce Tinder dating profiles on demand.
After thousands of conversations, I lifted the guardrails so high that AI Mark would only respond to key phrases or sentences, which makes it harder to distinguish him from a chatbot.
And what about my constituents? They were slightly more forgiving of my support for local business than social media was. Some have used AI Mark. He’s given answers on pothole repairs, streetlighting and housing concerns.
What didn’t it do? It didn’t save any time. I read every single transcript to ensure we didn’t miss any questions from constituents. I can see this technology working alongside a casework team, but it needs a lot of refinement. I took this leap to understand what AI might be capable of and what it isn’t yet. I understand why some dismissed the model out of hand, but I think the potential is real, even if that’s all it is for now – potential.
Until then, if you need help drafting a haiku or dating profile, you’ll have to come to me directly.
