AI isn't a computer. It's a bridge.

We get asked variations of this constantly. Is it safe to use AI for grant applications? Can we trust it with our annual report? Should we let it answer supporter emails? The question always carries an expectation of a binary answer. Safe or not safe. Yes or no.

That expectation makes sense if you think of AI as a computer in the traditional sense. Computers either do the calculation correctly or they don't. "Computer says no" is funny precisely because computers deal in absolutes.

AI doesn't work like that. A better mental model is a bridge.

When an engineer designs a bridge, they work with tolerances. Steel expands in heat and contracts in cold. The engineer doesn't know what the temperature will be every hour of every day for the next hundred years. They know rolling averages, historic extremes, the probability that temperatures in their location will stay between -5°C and +40°C. They design for those tolerances with margins built in.

There is always a possibility that the bridge you're driving across will fail because the engineer got the tolerances wrong. A minuscule possibility, but a real one. We accept this. We drive across bridges anyway, because the engineering is sound, the maintenance is ongoing, and the alternative would be far worse.

Brunel designed the Clifton Suspension Bridge. It was a major contributor to Bristol's economy because it connected an otherwise isolated part of the city, enabling trade that couldn't happen before. He was smart enough to think about the tolerances required, and since then other engineers have maintained and improved the design.

AI needs the same kind of thinking. Not "is it safe?" but "what are the failure modes, what are the tolerances, and are the mitigations proportionate to the stakes?"

Take the translation question. A large language model translating content can fail in several specific ways. It can mistranslate. It can get lazy and skip sections. It can overcomplicate the language. It might produce output that's programmatically corrupted. Those are the main failure modes for this task, and they increase or decrease with the complexity of what you're translating.

Then there's the question of stakes. Is someone reading this translation for a hobby? For their career? Or is it critically important information being delivered at a point when their life is in danger? The tolerance for error changes enormously.

You also have to weigh the counterfactual. A human translator has a base error rate too. They also take time, which means many things simply never get translated at all. No translation is also an outcome, and often a worse one than an imperfect translation.

Most of these failure modes can be engineered around. Translate smaller sections rather than entire documents. Simplify the source language before translation, then refine the output with domain-specific terminology afterwards. Add caveats so readers know the translation was AI-assisted. Run multiple automated translations and have a human select the best version. Guide the model with domain-specific instructions.

Our answer on stage was that automated translation is generally more valuable than the risks it creates, provided it's properly engineered. Not a snappy sales pitch. But the sales pitch for bridges would be equally long-winded if we properly engaged with the physics of how they're built.

Now apply the same framework to a different task: fully automating a charity's social media presence. An AI that plans the calendar, creates content, posts at optimal times, and responds to comments and direct messages. Run the tolerance analysis and the conclusion flips. The potential harm outweighs the benefit. An automated system would cut humans out of social listening entirely. You'd lose the ability to notice when a conversation shifts, when a safeguarding concern surfaces in a comment, or when your messaging is landing badly. The failure modes are higher impact and harder to mitigate.

Same framework. Different answer. That's the point.

The answer for charities is rarely going to be binary. It's trade-offs and engineered tolerances. But this doesn't mean doing nothing. To stretch the bridge metaphor one last time: it would be inconvenient to find yourself in London and reliant on ferry boats to cross the river. That was the reality not so long ago, and you can still see the economic consequences in east London, where bridges were slower to be built. There are real benefits to gain from AI, and declining to build anything carries its own costs.

The difference between bridges and AI is that we've had two to three thousand years of experience building bridges and watching them fall down. Most people's experience of AI is two to three years old, and we are very much still learning which designs hold and which don't.