AI capabilities is advancing rapidly, but it is very difficult to know for how much longer. It is also difficult to know in how many fields it will reach/surpass human expert level. AI is affecting basically all parts of knowledge work – or, rather, has the potential to do it: there’s a substantial capability overhang, where the current technology could be used to automate or augment current work, but isn’t used.
Even understanding the effects of current state of AI within a single sector is a big challenge: the most important effects are likely second- and third-order effects, most of which become visible only when the more direct effects are clear. Trying to predict how AI affects “knowledge, knowledge work and education” on a 5–10 year horizon, is basically impossible.
One might even argue that making a prediction about how AI affects knowledge work and education over 5–10 years is counterproductive: the possible outcomes are so different that preparing for a single future is likely to lead to decisions that put you in a worse position than doing at all. Locking on a single future is a bad strategy when you have radical uncertainty.
The solution, unsurprisingly, is to explore multiple scenarios and prepare for a variety of futures.
But before we go into details about this, let’s recap some things from the previous (and first) blog post.
This is a very brief (and human-made) summary of the first blog post in the series about how AI affects knowledge, knowledge work and education.
The series focuses on language-based AI, not generative AI or ML in general. This is motivated by LLMs being a general-purpose technology that advances and spreads rapidly. When I write “AI”, read it as “LLMs + scaffolding software”.
There are three dimensions that more important than others when it comes to forecasting impacts of AI:
How fast AI capability advances. We are currently seeing a nearly incomprehensible speed: from 2019 to 2025, AI has gone from barely counting to ten, to making meaningful contributions to scientific research.
How far AI capacity will advance. The “end state” for AI capabilities may vary for different fields, and we’ve seen LLMs match or surpass human experts in some of them.
How quickly AI will diffuse in society. We see rapid adoption of AI, but some knowledge becomes obsolete as AI advances, and adoption on a personal level doesn’t imply adoption at organizational or national levels.
I introduced a scale for describing AI capabilities in a given field. The most important steps are:
Level 2 – meaningful contribution: AI complements human experts by its general broad knowledge or its peaks of knowledge. Laymen can use AI as replacement for a fairly competent consultant and often get a good result.
Level 3 – fairly reliable performance: AI can perform tasks as well as an average person you could hire to do a job – including understanding and structuring tasks, improving on results, and asking for input or guidance when necessary. As with people you hire, you also get occasional mistakes and misunderstandings.
Level 4 – expert performance: AI performs most tasks in the field better than most human experts. Most mistakes are due to flawed instructions.
When faced with radical uncertainty – multiple futures are possible, with likelihoods that are difficult to assess and vastly different outcomes – we can reduce complexity and explore one possible scenario at a time.
Selecting how many scenarios to explore is a balance between available resources and how deeply you want to go into each scenario. When exploring how AI affects knowledge, knowledge work and education, I use the following perspectives for selecting scenarios:
Explore the edges. The three dimensions described in the previous blog post spans the game board. Some combinations of extremes can be excluded, the rest are candidates to explore.
Wanted/unwanted futures. This can use scenario end states as the starting point, working backwards to explain how they may appear. (Or discard them as implausible if no plausible chain can be found.)
Events with high likelihood. Starting in the present, some events and chains of events seem more likely than others. These can form scenarios, or seeds for scenarios.
Events with particularly high impact. Also, some low-probability events can have extreme impacts, making them worth exploring. They, too, can form scenarios or scenario seeds.
Some scenarios resulting from these perspectives will probably be the same, or very similar. For example, wanted/unwanted futures probably overlaps heavily with edge cases.
Using the perspectives above
AI slowdown, rapid diffusion (edge case): In this scenario, AI progress stalls and the current technology is utilised widely over a few years.
AI slowdown, slow diffusion (edge case): In this scenario, AI progress stalls, but efficient use of the technology is concentrated to a few for a long time.
Slow AI takeoff, fast diffusion (edge case): In this scenario, the ceiling for AI is high, but the technology advances at a manageable pace and spreads quickly.
Slow AI takeoff, slow diffusion (edge case): In this scenario, the ceiling for AI capacity is also high and advances at no faster rate than today, but use of the technology spreads slowly.
Fast AI takeoff (edge case): In this scenario, AI advances rapidly and surpasses human expert level in basically all cognitive work. Diffusion is slow in comparison. This scenario will be similar to #4 (slow AI takeoff, slow diffusion).
AI for augmentation, not replacement (wanted future): In this scenario, AI empowers humans to do more and become more. It is Star Trek rather than Wall-E.
AI for everyone (wanted future): In this scenario, the frontier AI and its benefits become available for everyone – all levels in society, every part of the world.
Fat, lazy and stupid (unwanted future): In this scenario, we gradually disempower ourselves by losing skills and competencies we rely on AI for. This is the “we are pets” future – the converse of #6 (AI for augmentation, not replacement).
AI for the elite (unwanted future): In this scenario, only a few have access to the most powerful AI, and its benefits are highly concentrated. It is the converse of #7 (AI for everyone).
AI dominates most knowledge work (likely path): In this scenario, AI reaches at least expert performance (level 4) in most knowledge work. Some niches remain unavailable for AI, as well as much outside knowledge work.
AI dominates most knowledge and physical labour (high impact path): In this scenario, AI reaches at least expert performance (level 4) in most knowledge work and physical work.
The US and China locks down AI access (high impact path): In this scenario, the current AI leaders decide not to share new generations of the technology with the rest of the world. It is a special case of #4 (slow AI takeoff, slow diffusion).
Analysing twelve scenarios is too much to do manually1, and there is significant overlap between the scenarios above. A curated list could look like this:
The imminent slowdown. This scenario is characterised by AI capabilities reaching a plateau within a year, which means that we have a fair chance of predicting AI capabilities by looking at what’s available today.
This scenario contains scenario 1 and 2 (AI slowdown, rapid/slow diffusion).
The manageable future. This scenario is characterised by a continued AI progress, but at a pace that is possible to adapt to (but also to fail adapting to).
Scenario 3, 4 and 10 (Slow takeoff, fast/slow diffusion and AI dominates knowledge work) are the core of this scenario.
Scenario 6 and 8 (AI for augmentation vs. fat, lazy and stupid), and 7 and 9 (AI for everyone vs AI for the elite) appear as variants.
Rupture. This scenario (or collection of scenarios) is characterised by big changes that are too rapid for the society to accomodate.
Scenario 5 and 11 (fast AI takeoff, slow diffusion, and AI dominates physical + cognitive labor) are the core of this scenario.
Scenario 12 (geopolitical lockdown) appear as a variant.
The scenarios are not predictions, but tools for thinking. The value comes from systematically exploring possible futures and synthesizing insights across them.
I won’t be able to do a rigorous analysis of the scenarios for these blog posts, but will still aim for this:
Trace second- and third-order effects as systems interact and adapt.
Identify bottlenecks and breaking points where systems fail or transform.
Look for pivotal moments where small decisions could significantly alter trajectories.
Examine who has agency to shape outcomes and how.
The results will be published in upcoming blog posts.
After examining each scenario individually, I'll look across them to identify:
Common patterns or inevitable developments that appear in multiple futures.
Critical uncertainties we should monitor as early warning signals.
Robust actions that are valuable in many different futures.
Ideally, scenarios should also evolve as reality provides new information. Were this commissioned scenario work rather than personal exploration, I’d revisit and update these scenarios periodically – assessing how likely they are, removing or adding new ones as warranted.
These scenarios are starting points for thinking about uncertain futures. If you see gaps, alternative trajectories, or critical factors I've missed, please share your thoughts in the comments.
In the next post, I'll dive into the Imminent Slowdown scenario – exploring what happens if AI capabilities plateau at roughly today's level.
But scenario analysis could be done using AI. I hope to write a blog post about this, too.
No posts
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.