“Prediction is difficult, especially about the future.”
– Yogi Berra (attributed)
Speculative scenarios have become an influential rhetorical device in dialog about AI. In April of 2025, AI 2027 made a major media splash, which essentially stated that artificial superintelligence will arrive in 2027 and revolutionize the world. In January of 2026, the authors pushed out their timelines. Another notable speculative piece, published in February of 2026 by Citrini Research, managed to cause a dip in the stock market. The piece forecasted economic instability due to advances in job automation via AI. In June of 2026, yet another long-form speculative piece took center stage, Europe 2031, which laid out a scenario of global unrest—caused by AI—that leaves Europe as a laggard.
These pieces are transparent about their use of speculation. However, the intention to influence policy—whether corporate or government—is also clear. Likewise, the works are not presented as science fiction meant to spur introspection—a separate category of literature with its own considerations. Consequently, reading this style of rhetoric with a critical eye is important. 1
Fast take-off. Oftentimes, speculative works will have some type of “how we got to where we are”. Pay close attention to cherry-picking of this history. First, these pieces often default to giving a short history, giving credence to the “fast takeoff” viewpoint. Rarely do they present the decades of progress that led to the current state of AI, including the long periods of plateaus. Second, these histories rarely mention AI failures or the lack of ROI being generated by LLM-driven systems. Third, beware of post-dictions, where the author positions something like a prediction when they have the benefit of hindsight (e.g., “it was obvious that, eventually, AI could write large volumes of software”).
Much speculation about AI rests on this “fast takeoff assumption”. That is, in the blink of an eye, AI models will progress beyond human intelligence and recursively self-improve. This perspective views AI progress as an exponential function rather than something like a logistic function (i.e., an exponential increase followed by a plateau). The viewpoint also does not give credence to the idea that LLM scaling laws simply provided a step change of performance and that more paradigm innovations are needed to unlock further major advancements.
Inevitability. In that vein, these pieces also often present novel breakthroughs as obvious or inevitable, despite the evidence found in research. Speculative articles often have an underpinning idea that we need massive large language models (LLMs), motivating the race to scale these models and build corresponding compute infrastructure. However, for the vast majority of tasks for which people currently use LLMs, they do not need or even benefit much from large models. The use of the largest, state-of-the-art models is often due to inertia and marketing, not necessity. Likewise, alternatives exist to LLM implementations (i.e., good, old-fashioned deterministic software). Speculative work is often based on the fact that we truly require massive models for a growing number of tasks.
Solvability. Like the assumption of inevitability, the underlying belief that “we can solve all problems” sustains AI speculation. For instance, actual continual learning (not just memory management via an external database) would assuredly be a lynchpin of “a fast takeoff”. Given current paradigms and performance, the null hypothesis is that we cannot robustly solve this problem with existing methodologies, but oftentimes we have a null hypothesis such as “this will be solved in a few years”. This position results in having the wrong baseline assumptions. Right now, we have not solved effective long-term memory for LLMs, so leaps to continual learning would be a true leap. The problem may eventually be solved, but we must be realistic about where we currently are in terms of progress.
Simplicity. AI speculation often fails to proactively grapple with the complexity of our world and the challenges of making even medium-range predictions. Much of AI speculative forecasting seems to be inspired by the science fiction concept of psychohistory, where human history can be predicted with mathematics (though modern AI speculative articles do not seem to employ even basic devices like Bayesian reasoning and confidence intervals). Informative forecasts about the future can be made when the scope is specific, humility is employed, and deep research is performed about the details (see the “superforecasters” from the Good Judgement Project as a prime example). AI speculation often fails to follow such protocols.
Predictability. AI speculation often presents predictions as point estimates (i.e., they present the best, the worst, or the average outcome). They often do not share a range of outcomes and the consequences of different outcomes. Effective predictions are not about the “happy path” but about calibrating uncertainty, grappling with entropy, and speaking to the interconnectedness of outcomes (i.e., if we are wrong about X, that also impacts the resulting chain of predictions). Speculation unusually displays little game theory thinking. The predictions are presented as being a closed system (if X happens, then Y happens). It does not give space for randomness or our known struggle to predict complex, multi-turn scenarios. A disclaimer of “this is speculation” at the beginning of the article is not sufficient. The branches of uncertainty should be recurrent, central elements.
Finally, details matter for prediction. For instance, if the detail of “we definitely need the best AI models for cybersecurity defense since the best models are used for offensive” is actually not true, a prediction chain predicated on this statement should also change. (And this statement is not, in fact, the best prediction of reality since old-fashioned cybersecurity can go a very long way. We don’t need AI to improve access controls, to prevent data exfiltration, to implement zero-trust security, or to sanitize input). If a detail is faulty, a whole chain of predictions can and should change. Errors cascade.
At the end of the day, “AI” is a vast enough topic to make essentially any case someone wants if they engage in cherry-picking of evidence and discounting of uncertainty. Speculative AI pieces are often too heavy on narrative and too light on substance and actual research. Believing arguments in speculative pieces can be worthwhile, but only if we have read and reflected with intention and a discerning eye.
1.To note, I authored a speculative AI novella in 2023. I agree that speculative works can be instructive as long as they are clearly positioned as fiction and/or philosophical inquiry.
Views and opinions expressed by authors and editors are their own and do not necessarily reflect the view of AI and Faith or any of its leadership.
