Microsoft’s Copilot AI tried its hand at predicting the 2026 NFL Draft first round—and promptly selected Arch Manning twice, assigned defensive tackle Peter Woods to three different teams, and ignored consensus top prospect Caleb Downs entirely. The experiment, published by USA TODAY on September 11, 2025, didn’t just produce a mock draft; it delivered a masterclass in why today’s conversational AI still can’t handle zero-sum games without deliberate guardrails.

The newsroom asked Copilot a deceptively simple question for each of the 32 picks: “With the __ pick in the 2026 NFL Draft, who will the [TEAM NAME] select?” Using the draft order from Tankathon after Week 1, the AI spit out a single-round mock that reads like a fever dream of football fandom. The results, while entertaining, expose fundamental limitations in how large language models reason about dynamic, constrained environments.

Duplicate Picks and Impossible Scenarios

The most glaring failure mode: Copilot repeatedly ignored the cardinal rule of draft logic—once a player is taken, they’re off the board. Arch Manning, the Texas quarterback and nephew of Peyton and Eli, went second overall to the New England Patriots… and then again at fourth overall to the Miami Dolphins. Peter Woods, a Clemson defensive tackle, landed with the Giants at No. 9, the Titans at No. 11, and the Raiders at No. 22. Linebacker Suntarine Perkins was picked by the Ravens at No. 6 and the Lions at No. 10. Cornerback Jermod McCoy heard his name called by the Rams at No. 7 and the Colts at No. 30.

These aren’t just amusing glitches; they reveal a structural deficiency. Because Copilot processed each pick as an isolated prompt, it had no memory of previous selections. It never built—or was never instructed to maintain—a live draft board. The assistant treated every query as a fresh start, blithely unaware that its own prior outputs should constrain future answers.

Questionable Positional Priorities

Beyond the duplicates, Copilot’s positional choices often defied common sense. The Baltimore Ravens, picking sixth overall—which would imply a catastrophic season—opted for linebacker Suntarine Perkins, a player widely projected for the second round. The Detroit Lions, likewise picking in the top 10, took the same Perkins, ignoring far greater needs. The New York Jets, with Justin Fields coming off a strong Week 1, still grabbed quarterback Cade Klubnik at No. 3. The New England Patriots, after investing the third overall pick in Drake Maye just two years earlier, doubled down on a QB with Manning.

These picks suggest Copilot leaned heavily on entrenched heuristics: quarterbacks are valuable, pass rushers are premium, and familiar names from recent headlines get recycled. It rewarded positional scarcity without accounting for existing roster investments, contract situations, or even the implausibility of certain franchise moves.

The Technical Underpinnings of AI Draft Chaos

Why did Copilot stumble so dramatically? The answers lie in the architecture and prompting approach.

1. Transactional Amnesia
The assistant answered per-team prompts in separate sessions, with no mechanism to enforce a draft-market constraint. Without a shared state, it couldn’t track that Arch Manning was already spoken for. This isn’t a hallucination in the traditional sense; it’s a mis-specified task. The model performed exactly as a stateless chatbot would—it gave the most plausible-sounding answer for each prompt, independent of all others.

2. Stale Data and Retrieval Latency
Conversational AI systems combine a reasoning core with retrieval layers that index news and databases. If the retrieval index lacks up-to-the-minute depth charts, injury reports, or scouting revisions, the model defaults to older priors. USA TODAY’s editorial team had to re-prompt Copilot multiple times when roster errors surfaced, a necessary human-in-the-loop step that most casual users would skip.

3. Heuristic-Heavy Scoring
Copilot frequently defaulted to prototypical outputs. In game predictions, it consistently forecast winners scoring in the mid-to-high 20s. In draft mode, it replayed a handful of recognizable names (Manning, Klubnik, Woods) that fit broad positional templates. The model favored “safe”, high-profile prospects rather than simulating the nuanced, fit-based calculus of real draft rooms.

4. Deterministic Outputs in a Probabilistic Domain
The NFL Draft is inherently probabilistic. Prospect grades, team schemes, medical red flags, and draft-day trades create a complex probability landscape. Yet Copilot returned single, deterministic picks—presenting each as a confident declaration rather than a distribution of possibilities. That framing hides the uncertainty and misleads readers into treating the mock as a forecast rather than an educated guess.

The Editorial Workflow: What USA TODAY Did Right

Despite the oddities, the newsroom’s broader Copilot project demonstrated a responsible blueprint for AI-assisted journalism. The team employed repeatable prompts, human verification steps, and selective re-prompting when the assistant leaned on stale data. For parallel experiment—game predictions—they cross-checked Patrick Mahomes’ Week 1 career totals against independent databases, confirmed Chargers roster moves via team and league reporting, and flagged any number that lacked a primary source.

That validation hygiene is the practical backbone of safe AI usage. When a pick rests on high-leverage facts (injuries, suspensions, trades), the rule is simple: withhold publication until an authoritative, independent confirmation exists. For the mock draft, however, the stateless prompt design undermined the entire exercise before verification could even begin.

Recommendations: Building a Smarter AI Draft Assistant

If newsrooms, fan sites, or teams want to use AI for responsible draft coverage, they must adopt several technical and procedural safeguards. Based on Copilot’s performance, here is a pragmatic playbook:

  • Enforce a Draft-Market State Engine: Run a single session that tracks selections and removes prospects from the available pool as picks are emitted. This prevents duplicates and forces a realistic zero-sum simulation.
  • Integrate Live, Authoritative Data Feeds: Ingest up-to-date depth charts, official injury reports, combine measurables, and beat reporter transactions to ground the model in reality.
  • Ask for Distributions, Not Single Picks: Prompt for probability bands (e.g., “Top three most likely selections with percentages”) and best/worst/most-likely scenarios. This conveys the inherent uncertainty and discourages overconfidence.
  • Maintain Human-in-the-Loop Verification: Every roster, injury, or contract claim must be validated against an independent official source before publication. Automate fact-checking flags to speed this process.
  • Disclose Model Identity and Data Cutoff: Always publish the assistant’s data cutoff timestamp and whether human edits were applied. Transparency builds trust and helps readers calibrate expectations.
  • Use Ensembling and Constrained Sampling: Combine multiple model runs or blend outputs from historic draft models, analytics services, and human scouts to average out stylistic biases.
  • Add Provenance Metadata: Track prompt variants, retrieval snapshots, and editorial corrections so each pick is auditable. accountability requires an unbroken chain of decision-making.

Where AI Actually Shines

For all the bungled picks, Copilot’s mock draft wasn’t worthless. The assistant excelled at speed (a full first round in seconds), explainability (it can instantly justify why it chose a player), and pattern surfacing (it quickly highlighted quarterbacks, pass rushers, and tackles as high-leverage positions). For a human analyst, these capabilities are a powerful research accelerator. The AI can run “what-if” scenarios, surface conventional wisdom, and generate draft copy that editors can refine.

USA TODAY’s experiment also proved that Copilot can revise its picks when given corrected facts—an iterative capability that supports dynamic editorial workflows. The core lesson: AI is ready to be a drafting-room assistant, not a standalone expert.

Risks Beyond the Mock Draft

Publishing AI-generated mock drafts carries risks that extend beyond embarrassment. Widely distributed deterministic picks can influence betting lines and fan expectations. If audiences treat Copilot’s choices as authoritative, the resulting market reactions can create self-fulfilling distortions. Factual errors—like relying on an outdated injury report or misstating a player’s eligibility—can damage trust and even invite legal exposure.

There’s also the danger of overconfidence. A single-answer output gives a false impression of certainty where humility is warranted. Readers deserve probability bands, clear data cutoffs, and transparent disclosure of the model’s limitations. For teams or leagues embedding AI into scouting or sideline workflows, governance becomes critical: audit trails, vendor lock-in concerns, and rigorous human oversight must be baked in from day one.

What This Means for Fans, Bettors, and Teams

  • Fans: Enjoy AI-powered mocks as conversation starters. They’re fast, fun, and often highlight player names worth watching—but they’re not definitive forecasts.
  • Bettors: Never treat a single deterministic pick as betting advice. Convert AI outputs into probabilistic signals, then triangulate with market odds and official reports.
  • Teams and Scouts: AI can accelerate scenario planning and tape analysis, but any operational use on draft day demands auditable provenance and layered human checks. The final call must always belong to experienced evaluators.

The Bottom Line

USA TODAY’s Copilot mock draft is a revealing case study, not a failure. It highlights the speed and pattern-recognition strengths of conversational AI while laying bare the design gaps that produce comical or misleading results. The duplicates, positional oddities, and roster misreads aren’t evidence that AI can’t do this work—they’re evidence that today’s models require market modeling, fresh authoritative inputs, and probabilistic framing to function in zero-sum systems.

For publishers, the path forward is clear: treat AI as a powerful assistant, not a replacement. Enforce draft-state constraints, integrate live data, demand distributions over single picks, and never skip the human verification step. Heed those safeguards, and Copilot-style tools can enrich draft coverage without sacrificing accuracy. Ignore them, and the next mock draft will still be entertaining—but nobody will trust it.