5 AI predictions for 2026

Datasparq's Head of Engineering shares his view of where AI is heading in 2026. To read more of Matt's thoughts, check out his personal blog on Medium

‍

I’ve noticed I’m quite good at predicting the future when it comes to AI. I’ve been doing nothing but solving real world problems with AI since 2019, which keeps me grounded in the reality of tech and immune to hype.

Last year I correctly predicted that Google would overtake their competitors in model performance, that video models would achieve Hollywood film level quality, and that specialist coding models would replace Stack Overflow, but would not replace human developers (yet).

1. AI models will finally learn to use a mouse quickly and accurately, leading to a revolution in automation and productivity

Agents can already read the screen and move the mouse today, but they’re painfully slow and inaccurate. This won’t last.

Once this is solved, users will be more and more tempted to give AI full access to their browsers, and then their whole computer.

Web browsers will be updated to encourage this, but with some safety controls. There will be a trend of designing sites that can be easily operated by AI.

Operating systems won’t be updated with any safety controls for AI operators for a while.

As a result of this, there will be more demand for ‘AI safety assurance’ within large businesses and governments — stuff like guardrails, permission models, and kill switches. We probably won’t see any updates to information security standards (ISO27001) to cover these risks for another few years.

2. Real time streaming models, and larger context windows for audio and video, will change how people work

It’s currently too expensive to feed a recording of your entire work day into AI for analysis, but imagine if it wasn’t. Transcripts are only so useful—they’re missing the facial expressions, emphasis, tone, etc.

We need to anticipate future AI models that can handle this much data as an input.

Businesses will start requiring their staff to record everything: all meetings (video and audio) + their screens as they work. This will allow AI Agents to truly understand how we work, learn the nuance of our jobs, and become genuinely useful as a full time member of the team.

3. Continuous learning will be the big new thing, leading to businesses getting locked in with their LLM providers

The major players will start marketing ‘teachable’ and ‘personalised’ AI. Businesses will find themselves unable to switch model providers because they can’t afford to lose this personalisation, and they can’t just take it with them to another provider.

Open source frameworks will appear, allowing businesses to invest in building and hosting this capability for themselves without vendor lock-in.

4. There will be more high profile cyber attacks, powered by AI, and data leaks from vibe-coded shovelware apps

Nobody is safe.

5. Another open weights model will jump ahead of Google, Anthropic, and OpenAI

Currently, Google and Anthropic are far ahead of the competition, but I doubt they’re feeling complacent. Alibaba, Moonshot, and Z.ai are achieving incredible results while giving their models away for free.

We should expect to see many innovations in model speed, reasoning ability, and VRAM requirements from the wider AI research community (mainly from China), which will lead to another sudden drop in Nvidia’s stock price.

‍