The start of 2026 has produced an unexpected subplot to the Premier League season: the weekly predictions contest run by BBC Sport — fronted by pundit
Chris Sutton — is now being outscored by an artificial intelligence, with
Microsoft Copilot Chat topping the leaderboard and taking another weekly win while human forecasters and guest celebrities chase the numbers. Sutton’s good early-season form has stalled in the predictions title race, and for week 20 he faced singer‑songwriter and Newcastle United fan
Andrew Cushin, as the AI continued to underline the growing role of machine assistance in forecasting football results.
Background
What the BBC predictions series is and how it scores
BBC Sport’s weekly predictions feature asks a resident pundit — currently Chris Sutton — to submit exact scores for every Premier League fixture, then pits those forecasts against a weekly guest, BBC Sport readers and an AI that the editorial team asks to “predict this weekend’s Premier League scores.” The scoring is straightforward:
10 points for a correct result (win/draw/loss) and
40 points for an exact correct scoreline. That weighting strongly rewards correct exact-score calls while still valuing correct outcomes.
The format is designed for engagement: readers can enter their own predictions and compare with pundits, celebrities and the AI. Over a full 380‑match season this becomes a cumulative contest — a test of consistency and calibration rather than one‑off luck.
Recent context: Lawro to Sutton, and the AI entry
The predictions tradition stretches back years on BBC platforms with well‑known names (Mark “Lawro” Lawrenson being the most famous example). Sutton inherited the weekly correct‑score mantle in recent seasons and the segment remains a staple for fans who enjoy the mix of analysis, bragging rights and light entertainment. In the current season the editorial team has systematically added an AI competitor, using Copilot Chat as the on‑duty machine forecaster. That AI has not only matched human-level output in some weeks but has taken outright weekly wins by combining multiple correct results with the higher‑value exact‑score calls.
Week 20 snapshot: Sutton, Cushin and Copilot
For the fixtures in early January, Sutton’s column features his reasoning and a set of scorelines. His guest for the week — Andrew Cushin, an indie singer‑songwriter who has been building momentum since his second album release in May 2025 — provided competing picks and local insight on matches such as Newcastle at home. The BBC disclosed that the AI’s picks were generated by prompting Microsoft Copilot Chat with a simple instruction to predict the weekend’s scores. In week 19 the AI’s combination of correct results and two exact scores earned it a 100‑point weekly haul and the weekly victory.
This is notable for a couple of reasons. First, the editorial decision to include an AI that is asked a one‑line prompt — rather than a bespoke, sports‑trained model — raises questions over the
type of AI being compared to traditional punditry. Second, the AI’s success shows that even off‑the‑shelf conversational assistants can be useful in routine forecasting tasks when fed the right prompt and current data.
Andrew Cushin: a guest profile (and a reminder of football’s cultural reach)
Andrew Cushin’s appearance in the predictions column is more than celebrity name‑dropping. Cushin’s second album,
Love Is For Everyone, released in May 2025, reached the UK Top 40 and his touring schedule — supporting bigger names and staging sold‑out local shows — underlines the crossover between football fandom and cultural life. He is, by his own admission, a proud Newcastle United fan whose local knowledge and matchday instincts feed into his choices. That profile is a useful reminder that the BBC series blends quantitative forecasting with human stories and personalities — an editorial mix that drives clicks, debate and social engagement.
Why an AI can outperform pundits on a weekly leaderboard
It’s tempting to write this off as a novelty: an algorithm squeaks ahead for a week and the story fades. But the mechanics of football forecasting explain why even a general-purpose AI — when asked and calibrated in certain ways — can challenge human forecasters consistently.
- Scale of data processing: AI systems can synthesise vast, structured datasets (xG, recent form, head‑to‑head stats, injuries, fixture congestion) far faster than a human can. Machine models commonly used for match prediction are trained to detect patterns across thousands of matches and seasons. That gives AI an edge in spotting probabilistic edges that are easy to miss in narrative journalistic workflows.
- Consistency and calibration: Human pundits are prone to recency bias, emotional attachment to clubs and headline‑seeking selections. AI, given the same latest inputs, will produce predictions with consistent heuristics, avoiding emotional drift. That consistency can translate to better cumulative leaderboard performance even when the AI doesn’t “feel” right.
- Objective pattern recognition: Modern machine learning models — from gradient boosting to transformer‑based sequence models — identify non‑obvious correlations (for example, how playing style and xG differential interact with specific referees or weather) and exploit them. Academic surveys and papers show AI models routinely exceed baseline human accuracy on match outcomes, typically improving on random chance by a sizeable margin.
These are real technical advantages, and they translate into points under the BBC’s scoring system. A machine getting the result right frequently (10 points) and occasionally nailing the exact score (40 points) will compound leads over time.
Limits and blind spots: where humans still matter
That said, AI’s lead is far from a knockout blow. There are clear, material limitations to using a conversational assistant like Copilot Chat for live football forecasting — and these create niches where human expertise still outperforms or adds indispensable value.
- Last‑minute, non‑public information: Manager op‑outs, last‑minute squad choices, illness or dismissal at training, and tactical rotation intentions often leak through club communications channels and personal networks. Humans with trusted contacts or local knowledge can access and interpret those signals faster than an AI that is not explicitly fed that proprietary or ephemeral data.
- Contextual nuance and strategy: Pundits bring tactical frameworks and interpretive judgment. They can link coaching choices to entrenched football philosophies, simulate how a manager might react in a knockout sequence, and weigh psychological factors (e.g., revenge fixtures, cup fatigue). AI models are increasingly good at simulating these but still struggle with causal inference when the training data lacks comparable historical analogues.
- Black‑box and hallucination risks: Off‑the‑shelf conversational models can hallucinate or generate overconfident predictions without transparent probability estimates. When a BBC article says it “asked Copilot Chat” to predict scores, the editorial team is using a general assistant that doesn’t necessarily expose the confidence, underlying feature set or model assumptions. That opacity matters if readers and bettors treat AI outputs as authoritative.
- Randomness of sport: No model fully overcomes the inherent randomness in football — a freak deflection, a refereeing error, a moment of individual brilliance. Academic work stresses that even high‑performing models achieve accuracy in the 54–68% range depending on task framing and dataset; they improve odds but do not guarantee outcomes.
Editorial transparency: the BBC’s methodology and an ethical checklist
The BBC discloses the AI used and provides a human narrative alongside its picks, which is best practice for editorial transparency. However, there are areas where clarity would benefit readers and the broader sports journalism ecosystem:
- Model specifics: Identify whether the AI used is a general chat assistant or a sports‑trained predictive model. The editorial note that Copilot Chat was asked to “predict this weekend’s Premier League scores” is transparent in tool identity but vague on inputs and context. Readers deserve to know whether the AI used live injury feeds, official team sheets or only public, historical data.
- Confidence scores and probability distributions: Publishing probabilistic outputs (e.g., expected probabilities for home win/draw/away and likelihood of exact scores) would make AI predictions far more informative and comparable with betting markets and odds providers.
- Versioning and reproducibility: Conversational models get frequent backend updates. A small change in Copilot’s inference engine can alter outputs from week to week. Flagging the date and model version, or archiving the prompt and output, improves reproducibility.
- Disclosure of human intervention: If editors or pundits curate the AI outputs (e.g., override certain scores), it should be stated. That prevents the impression that the AI is infallible or purely automated when human judgment shaped the final published picks.
These steps would reduce the risk of overreliance on AI output and strengthen audience trust.
Broader implications: journalism, fandom and the betting ecosystem
The BBC feature sits at the intersection of journalism, entertainment and the betting ecosystem. The rise of AI predictions carries multiple consequences.
- For sports journalism: AI can augment reporting — quickly generating probabilistic forecasts, highlighting trends and surfacing unusual correlations. But editorial roles will shift from “sole expert forecaster” to “curator and explainer of model outputs,” with greater emphasis on interpretation, narrative and verification.
- For fans: The AI element creates a new engagement vector. Fans enjoy comparing opinions and measuring their intuition against an algorithm. That interaction can deepen engagement; it also risks commodifying punditry into a numbers game if not paired with strong storytelling.
- For bookmakers and odds markets: Bookmakers already run sophisticated models; a general‑purpose AI’s public forecasts are unlikely to beat market prices. But transparency about model inputs could influence in‑play betting and micro‑markets. Gamblers should treat the BBC’s AI as entertainment and, where stakes are involved, consult professional market odds and probit outputs rather than raw predictions. Academic literature shows models often improve on naïve human baselines but still fall short of bookmakers that incorporate market information.
- For AI product design: The Copilot experiment suggests demand for specialized sports‑prediction APIs that expose probabilities, explainable features and audit trails. Vendors and publishers could partner to build purpose‑built tools with documented data pipelines.
Risks and best practices for publishers
The introduction of AI into public forecasting is grey with opportunity and risk. Media organisations should adopt guardrails to preserve credibility:
- Never pass off AI as infallible. Make sure every AI prediction page includes clear contextual language about probabilistic nature and historical performance.
- Publish performance metrics. Weekly leaderboards should include win rates, exact‑score hit rates and trend lines so readers can see how humans and machines compare over time.
- Audit models and prompts. Keep an internal log of the prompts used and evaluate AI outputs against ground truth post‑match. If a conversational assistant is used, include a short description of preprocessing (e.g., “we provided current league table, last five matches and injury lists to the model”).
- Protect readers from gambling harms. Given the easy link from predictions pages to betting markets, include links to responsible gambling resources and discourage chasing losses.
These measures protect audiences and preserve the editorial value of punditry.
How readers should treat the weekly leaderboard
For readers who enjoy the weekly predictions as a pastime, the leaderboard is a simple scoreboard. But for anyone thinking about model outputs as advice for wagering or forecasting, treat the results as one input among many. Consider the following practical approach:
- Check the model’s recent accuracy on match outcomes and exact scores.
- Compare AI probabilities with bookmaker odds to detect market discrepancies.
- Factor in non‑public variables (confirmed team news, weather, travel) before acting.
- Use probability thresholds (for example, only consider AI suggestions where the home win probability exceeds 60%) rather than raw scorelines.
This method converts a click‑bait prediction into a disciplined, probabilistic decision framework.
The human angle: Sutton’s candid self‑assessment and the limits of ego
Part of the charm and value of the BBC predictions series is its human texture. Chris Sutton’s public acknowledgement that “it’s been a bad start to the new year for me, with AI top of the table” and his offhand remark about losing to his daughter at cards blend humility with personality. That human voice keeps the feature readable and relatable, and it’s precisely what a machine cannot replace: the lived sense of fandom, humor and lived expertise.
That blend matters. Even if AI continues to score more weekly points over a season, the BBC’s editorial product is not only a raw leaderboard; it’s a conversation piece that stitches together analysis, anecdote and cultural context. Pundits who can interpret model outputs, add anecdotal evidence and tell the story behind a pick will retain value.
Conclusion: augmentation, not annihilation
The early 2026 leaderboard — with Copilot Chat slipping ahead of Chris Sutton and celebrity guests like Andrew Cushin — is an instructive case study in how AI is reshaping routine forecasting tasks in sports media. The AI’s advantage is rooted in scale, pattern recognition and consistency; its limitations are transparency, last‑mile contextual awareness and susceptibility to randomness.
The sensible editorial posture is augmentation: use AI to surface probabilities, crunch numbers and flag anomalies, while preserving human judgment to interpret, contextualise and narrate. Audiences benefit when outlets publish both the model outputs and the editorial reasoning that accompanies them. That dual approach upholds journalistic standards while embracing technological progress.
In short: AI will continue to haunt the Premier League predictions leaderboard, but it is the editors, pundits and storytelling that will keep readers coming back — and that human‑machine partnership is where the real future of sports journalism lies.
Source: BBC
Premier League predictions: Chris Sutton v singer-songwriter Andrew Cushin - and AI