Many areas of artificial intelligence, and machine learning in particular, aim at being probably correct, i.e., valid on average, rather than pursuing the idealistic goal of being provably valid for all inputs. However, AI systems could still be predictably valid, such as an imperfect robot deliverer for which we can reliably and precisely predict the task instances for which it is correct and safe, its valid operating range. “Predictable AI” is a nascent research area that explores ways of anticipating key validity indicators (e.g., performance, safety) of present and future AI ecosystems. We argue that achieving predictability is crucial for fostering trust, liability, control, alignment and safety of AI, and thus should be prioritised over performance. We formally characterise predictability, explore its most relevant components, illustrate what can be predicted, describe alternative candidates for predictors, as well as the trade-offs between maximising validity and predictability. To illustrate these concepts, we bring an array of illustrative examples covering diverse ecosystem configurations. “Predictable AI” is related to other areas of technical and non-technical AI research, but have distinctive questions, hypotheses, techniques and challenges. This paper aims to elucidate them, calls for identifying paths towards a landscape of predictably valid AI systems and outlines the potential impact of this emergent field.

Contamination in large language models (LLMs), and machine learning more broadly, refers to the inclusion of equal --or very similar-- examples in both training and test sets. This phenomenon usually translates into better test performance. Here we explore when this contamination is performed intentionally, for purposes that can be malicious (e.g., get better scores in evaluations) or benevolent (e.g., fix some mistakes). These interventions, usually in the form of fine-tuning memorisations, come with a budget in the size of the fine-tuning dataset. Several trade-offs appear between the breadth of the intervention (how many examples to be memorised), its depth (how many repetitions of each example) and the difficulty of the examples. By studying several LLMs and datasets, we observe some monotonic behaviour (more difficult items require more depth to be `fixed') but also some non-monotonic phenomena (very high depth levels have negative effects on non-contaminated examples). This suggests that trade-offs should be found not only in terms of the budget but also according to model specifics, the task and the item difficulty at hand.

The proportion of AI-generated content permeating the well of knowledge is increasing significantly. Large language models (LLMs) contribute to that contamination but they also suffer from it. However, it is yet to be clarified the effect of different sources of error, be it human-generated or LLM-generated. Controlling for the percentage of error, we explore the impact on LLM fine-tuning when errors come from humans, from other language models or are generated randomly using an aleatoric or epistemic source. In this paper, we compare these different types of error for in-distribution and out-of-distribution experimental settings. By analysing the levels of errors and their distribution, we find a nuanced view: while in-distribution human-generated noise seems more benign than the LLM-generated counterpart, in the out-of-distribution case the model-generated noise may not be necessarily worse.

Adversarial benchmark construction, where harder instances challenge new generations of AI systems, is becoming the norm. While this approach may lead to better machine learning models —on average and for the new benchmark—, it is unclear how these models behave on the original distribution. Two opposing effects are intertwined here. On the one hand, the adversarial benchmark has a higher proportion of difficult instances, with lower expected performance. On the other hand, models trained on the adversarial benchmark may improve on these difficult instances (but may also neglect some easy ones). To disentangle these two effects we can control for difficulty, showing that we can recover the performance on the original distribution, provided the harder instances were obtained from this distribution in the first place. We show this difficulty-aware rectification works in practice, through a series of experiments with several benchmark construction schemas and the use of a populational difficulty metric. As a take-away message, instead of distributional averages we recommend using difficulty-conditioned characteristic curves when evaluating models built with adversarial benchmarks.