AI Peyton Manning – Can AI Read a Defense Before the Snap?

1. Want to Get the Most Out of Watching NFL? Start With Prime Video

Quick question for you NFL fans: do you have a go-to broadcast preference?

Personally, I'm not all that particular — FOX or CBS, whoever's in the booth, it doesn't make or break things for me (though getting Drew Brees is always a plus, even if his analyst game still has room to grow). But when the game I want to watch falls on Thursday Night Football on Amazon? That's a lucky week. The stream quality on Prime Video is noticeably better, the delay is minimal — but the real draw, honestly, is Prime Vision, Amazon's player-tracking-powered alternate broadcast experience.

1.1. What Is Prime Vision?

Ever since Amazon locked down the exclusive streaming rights to TNF, they've built out an alternate feed that layers Next Gen Stats data — routes, completion probability, and more — over a camera angle roughly equivalent to All-22 film height.

Pass routes displayed on screen, from Amazon's Prime Vision introduction site

It used to feel like playing Madden while watching a real game. But lately, Amazon has been rolling out a wave of new AI capabilities under the "Prime Insights" banner. As TNF producer Sam Schwartzstein puts it, the goal is to let viewers "watch the game like a 10-year veteran" — and as someone who has never lined up at any level, having these tools genuinely raises my tactical understanding and makes the game richer to watch. Here are a few examples:

1.2. Defensive Alerts (Blitz Prediction)

Check out this clip from the 2023 TNF matchup between JAX and NO. The one where Foster Moreau dropped a TD pass in the end and they lost. Before the snap, players tagged with a red circle are the ones the AI has flagged as likely blitz threats — linebackers or defensive backs coming on a pass rush. You can see it first highlights what looks like an LB in front of TE Hill (#7), then picks up the CB who's just set up in front of slot WR Thomas (#13).

What makes this especially impressive is something the aforementioned Sam Schwartzstein mentioned when he appeared on The Athletic's podcast: Sean Payton told him, "It mostly lines up with what any football guy would tell you — but sometimes I'd see a player highlighted and have no idea why. Then that guy would actually blitz. It changed the way I look at blitzes."

And sure enough, looking at that same clip — Derek Carr clearly didn't read it as cleanly as the AI did, because he ends up taking pressure and throwing an incompletion.

You'll also notice a green circle pop up on slot receiver Michael Thomas at the end of the clip. That's the Prime Targets feature, which highlights open receivers who are in position to pick up a first down.

1.3. Coverage ID (Coverage Prediction)

Next: coverage prediction. This is closely tied to the research paper I'll be getting into later — the concept is that the system reads defensive positioning, alignment, and movement to predict whether the coverage is man or zone before the snap. The clip below from the 2024 DEN@NO game makes it crystal clear. Yes, that's the one where Sean Payton came back to New Orleans and systematically dismantled his old team.

At the 0:02 mark, the bottom-right of the screen reads "man coverage." And sure enough, CB #1 Tailor is shadowing WR #84 Humphrey step for step and bats the pass away. I'll be honest — even I could sort of sense man coverage from how tight the corners were aligned — but to have it confirmed algorithmically in real time, purely from pre-snap positioning and movement, is still pretty remarkable.

1.4. Pocket Health (QB Pressure Monitoring — New in 2025)

This one rolled out in 2025, and it does exactly what it sounds like: a live visualization of the pressure the QB is facing in the pocket from the pass rush (green = clean, red = danger). The clip is from this year's BUF@HOU — Josh Allen's interception play. You can watch the pocket light up red from Houston's D-line push, Allen feeling it and firing on the run.

There are obvious downstream applications here, too: you could use this data to evaluate OL pass protection or measure a QB's pocket escape ability more objectively — so you wouldn't have to rely entirely on a few guys watching film and grading by hand.

1.5. Four-Down Territory

Go for it on fourth!

This one actually showed up in the first clip as well. On 3rd down, a blue line on the field marks the cutoff: if you push past this point, the AI recommends going for it on 4th down. I can already hear some fans yelling that the line is drawn way too conservatively — "we should be going for it from way further back than that!"

1.6. How Prime Vision Actually Works

So when I say these predictions are "AI-powered," what's actually happening under the hood?

• Every player's shoulder pads, the football, referees, the down marker sticks, and the pylons have 2–3 RFID chips embedded in them
• 20–30 receivers positioned around the stadium pick up those chip signals 10 times per second
• Those receivers triangulate "this chip is at this exact spot in the stadium right now," and Amazon's AWS pipeline cleans that into coordinate data — that's what flows out as Next Gen Stats (with positional accuracy down to a few inches)

The result: every play from snap to catch to tackle is recorded as a sequence of X-Y coordinates over time — i.e., tracking data. Prime Vision's AI features use those coordinates to estimate in real time things like how tightly defenders are converging, how open each receiver is, and whether a pre-snap alignment matches historical blitz patterns — and then overlay all of it on the live broadcast.

2. The Pre-Snap Read: An Information War

The AI stuff is genuinely impressive — but what makes it especially interesting is that so much of the most valuable information is available before the snap even happens.

And this isn't just about TV production. For QBs and coaches, it's everything. The offense is trying to decode the defense and find a crack to exploit in the run or pass game. The defense is working just as hard to make sure nothing is readable. The two main tools in this arms race have been exploding across the modern NFL:

• Shift/Motion: Pre-snap movement by the offense — both to disguise what play is coming and to read how the defense responds, gaining information about their coverage
• Disguise: Defensive movement designed to show one coverage look before the snap and rotate into something completely different after it, leaving the QB second-guessing himself

2.1. Shift/Motion

Motion started out as a coverage diagnostic — if a WR goes in motion and a defender follows him, you can reasonably guess man coverage. But the modern NFL has stretched that concept into entirely new territory: shifting numbers advantages, deploying versatile weapons in multiple roles, you name it. The Shanahan/McVay coaching tree is synonymous with motion-heavy offenses, and at this point, not using pre-snap motion is essentially malpractice. Per Next Gen Stats, the rate of plays featuring motion at the snap was just 4% back in 2017 — by 2023 it was 22%, and by 2024 it had jumped to 28%.

The hottest current example? Ben Johnson deploying jet motion and pulling linemen with the Chicago Bears.

2.2. Disguise

Of course, the defense isn't just going to sit back and absorb it. They fight back with disguise — showing one coverage pre-snap and spinning into something entirely different post-snap, pulling the rug out from under the QB. Rather than trying to explain "rotating from 2-high into Cover 3..." in text — which never lands right — I'll let coverage specialist DB Island do the talking:

Being a starting QB in today's NFL is genuinely brutal. You can't just read safety depth. You can't just check whether a defender follows motion. A QB is processing something like: "Looks like 2-high... wait, the free safety just crept up a step. Post rotation? Cover 1 Robber? But the nickel didn't trail the motion — he passed off to the outside corner. Zone-ish? But the Mike is shaded a bit outside, front looks like a 5-man..." — and they're doing all of that while simultaneously checking protection, calling an audible if needed, directing motion, and then actually delivering an accurate ball while avoiding pressure.

The volume of information a QB has to process in the seconds before and after the snap is almost incomprehensible.

Peyton Manning is the gold standard for reading a defense at the line, but every great QB in the modern era has had to get sharper at this as the game has evolved. Raw athleticism obviously matters, but you genuinely cannot play quarterback at the highest level without serious football intelligence. Mahomes, Allen, Burrow, Lamar — watch any of them closely and you can actually see the processing speed in real time.

3. Today's Research: Teaching an AI to Read Coverage Before the Snap

That was a longer setup than usual — but it's the necessary backdrop for the research I want to break down. The NFL Big Data Bowl is an annual data competition run by the NFL, where competitors work with tracking data from Next Gen Stats and other sources on a given theme. (For background on how the Big Data Bowl works, check out last year's piece on tackle metrics.) The 2025 theme was pre-snap and post-snap prediction.

The winning submission was a study called "Exposing Coverage Tells in the Presnap" — and the headline finding is: they built an AI that correctly identifies man vs. zone coverage with 90% accuracy before the snap.

3.1. Building and Training the Model

At its core, this is a machine learning problem. The researchers extracted all passing plays from Weeks 1–8 of the 2022 NFL season, then trained a model on pairs of data: (frame-by-frame positional information before and after the snap) + (whether that play was man or zone coverage). There was some sophisticated data engineering under the hood — I'll spare you the details, though the figure below gives you a sense of the methodology.

When they validated the model against Week 9 data — giving it only positional coordinates and asking it to identify the coverage — the results were striking: 89% accuracy at the moment of the snap, climbing to 93% within 1 second after the snap.

Source: cited paper above

3.2. An AI That Won't Get Fooled by Motion

Now, Amazon's Coverage ID feature reportedly hits around 95% accuracy on a much larger dataset — so the raw numbers alone aren't the story here. What makes this research genuinely interesting is that the authors open the black box and show exactly when and why the model makes its calls.

Take this play from CAR@CIN (2022 Week 9). Top-left is the live broadcast; right side is the tracking data visualization; bottom-left is the model's output (x-axis = time, y-axis = probability the coverage is zone).

Three things to watch:

1. Right away, when CAR WR #12 Shi Smith goes in motion and CIN nickel #35 Jalen Davis follows him, the model reads this as evidence of man coverage — and the zone probability drops sharply on the graph.
2. Then, at 0:03, #35 stops short of where Smith settles, and the linebackers shift. The model recalibrates: it reads this as disguise, and the zone probability surges back to 90%.
3. At 0:05, S #30 Jessie Bates shifts deep, and the model locks in: zone probability hits 98%.

The AI is picking up on the same subtle cues a veteran defensive coordinator would recognize — movement nuances that are genuinely easy to miss if you don't know what to look for. (Honestly, I only caught step 1 on my first watch.)

3.3. Film Study Applications: Know Which Defender to Key On

Here's where it gets really valuable. The authors argue that the model can work as a film study tool — helping QBs and coaches identify which specific defenders are giving away the coverage.

The next clip is from the LAC@JAX playoff game in 2022–23 — yes, the comeback that Chargers fans are still trying to forget — showing Trevor Lawrence getting burned by a well-crafted LAC disguise.

• From 0:01–0:04, LAC S #24 Nasir Adderley creeps toward the line while #32 Alohi Gilman drops back, manufacturing the look of Cover 1 Robber (man coverage).
• Then, right before the snap, Adderley (#24) retreats to his deep zone, exposing the actual Cover 2 (zone coverage).
• Lawrence doesn't catch the switch, throws into the zone, and CB #26 Asante Samuel Jr. makes the interception.
  

Pause the clip at 0:05. The model shows that Adderley's final direction is the critical tell: step forward and it's man, step back and it's zone. If Lawrence had studied this coverage tendency beforehand — "whenever this look comes out, just track the deep safety's last move before the snap" — he might have made a different read.

The paper includes a second example: Baker Mayfield getting picked off by a disguised Denver defense in the 2024 DEN@TB game.

• DEN's defense shows Cover 1 body language, with near-side S #22 Brandon Jones as the only deep dropper — and with CB #2 Patrick Surtain II pressing WR Mike Evans, it looks every bit like Surtain is locked into man coverage
• Mayfield reads the 1-on-1 for Evans and pulls the trigger — but the actual call is a disguised Cover 2, and S #22 Brandon Jones is sitting right there for the interception
• The model, however, had already identified zone — because CB #29 Ja'Quan McMillan was lined up 5 yards off and shaded outside rather than pressing, a classic zone tell
  

In the first example, it's Adderley; in the second, it's McMillan. Inside every disguise scheme, there's usually a player who inadvertently gives it away — what the paper calls a "giveaway" player. Identify who that player is in film study, and you've got a real edge.

3.4. Team-by-Team Breakdown

The analysis goes even further. The paper breaks down results by team to identify which specific defenders are the most consistent coverage tells for each unit. The researchers isolate instances where the model's confidence shifted 10%+ within 0.5 seconds — moments where a specific movement made the coverage obvious — then analyze which players were moving the most during those windows.

The outputs are remarkably actionable: for a team like the Jets, the two safeties both shifting to deep alignment is a strong zone signal; for Miami, CB Howard's burst out of press coverage tips the hand on what the disguise is doing. Exactly the kind of pre-game intelligence an offensive coordinator would kill for.

Apparently a NE-centric breakdown of AFC East defenses — guessing the authors might be Patriots fans

And that's without motion in the picture. The paper also shows how you can run the same analysis on motion plays — tracking how defensive movement patterns differ between man and zone reactions, and identifying the key tells for each team.

A heat map showing how Panthers defenders respond to motion. The kind of pre-snap intelligence that would be enormously valuable to have before kickoff

3.5. Oh, and the Authors Were Poached by the Eagles

At this point you're probably wondering: who are these people? Brace yourself — they were two undergraduates.

Yes, they had institutional support along the way. But still. That's ridiculous.

Your first thought is probably, "Okay, a team should just hire them outright." Turns out someone already had that same thought: two months after winning the Big Data Bowl, lead author Smit Bajaj was hired by the Philadelphia Eagles. Classic Howie Roseman. ...although when your offensive coordinator is who it is, being able to read coverages doesn't necessarily translate to what happens on the field. It's not exactly showing up in the results.

4. Takeaways

So — to put a bow on it: the combination of player tracking data and machine learning is producing genuinely transformative tools for understanding football. On the viewer-facing side, Prime Vision on TNF makes the game more legible than it's ever been for the average fan. Underneath it, the Big Data Bowl gives researchers a platform to crack open those black-box models and show exactly what's driving the predictions. And the coaching staffs and front offices that figure out how to put these tools to work will have tangible advantages over the ones that don't.
One more thing before we wrap: the 2026 Big Data Bowl theme has been announced as "predicting pass plays using tracking data." The finalist submissions will absolutely be worth your time — I'll plan to break down the winners here again next year.

Thanks for sticking with a long one. Appreciate the read.

5. Resources & Recommended Follows

Amefutobu (@AmefootClub)
"Amefutobu" is a Japanese word meaning a club of american football. When you google this Japanese word in Google, he ranks above actual college football program websites in Japan — which tells you something. The fundamentals breakdowns are genuinely excellent, and the coach-by-coach scheme write-ups are the kind of thing you bookmark and come back to. Always a good read, on the site and on X. (Diehard Ole Miss fan in the profile, but that avatar color palette is unmistakably Saints energy.)

Makotomoya (DB Island) (@FitzNFL)

Probably the best coverage expert of pass coverage in Japan. His timeline is a steady feed of coverage and disguise film breakdowns — the kind of stuff that makes you actually understand what's happening in the secondary. A ranking of the best DBs, scheme breakdowns by coordinator, you name it. The magnum opus: a classification of 150+ NFL pass coverage types, which is either the most impressive thing you'll ever come across or a sign of a deeply concerning obsession. Probably both. Also excellent on analytics and advanced metrics.

Prime Vision / TNF References

• https://www.aboutamazon.com/news/entertainment/thursday-night-football-amazon-prime-video-ai-features — The best single resource walking through each Prime Insights feature in detail.
  
• Sam Schwartzstein (@schwartzsteins) leads the project — worth a follow if you want to stay ahead of what's coming next.