Harvard's AI Is Hunting UFOs

Harvard’s new UFO hunt starts with a mundane reality: our skies are jam-packed with ordinary stuff. Every clear night, cameras and telescopes drown in streaks from satellites, blinking aircraft, drifting balloons, and atmospheric clutter that turns raw data into a visual traffic jam.

Layer in the daytime chaos—flocks of birds, helicopters hugging city skylines, private drones skimming backyards—and the signal-to-noise problem explodes. Any system staring up for hours will see millions of frames dominated by familiar, boring objects.

For decades, that background has sabotaged serious anomaly detection. Human analysts burn out sifting grainy footage, mislabeling Venus as a UFO or ignoring a fast-moving speck that doesn’t match any cataloged orbit. Most “unidentified” reports die in the gray zone between bad sensors and overstretched attention spans.

Modern sky surveys only amplify the problem. High-speed all-sky cameras, radar networks, and infrared sensors generate terabytes of imagery, while Starlink and other mega-constellations add thousands of reflective points racing overhead. The haystack keeps growing; the hypothetical needle stays the same size.

Harvard astrophysicist Avi Loeb and the Galileo Project frame it bluntly: we know almost everything we usually see. Birds, airplanes, satellites, helicopters, and natural phenomena like meteors and lenticular clouds account for the overwhelming majority of detections, across every time of day and illumination level.

That familiarity creates an opportunity. If you can rigorously teach Machine Learning systems what “normal” looks like—at different sun angles, weather conditions, and viewing geometries—then anything that refuses to fit becomes mathematically interesting. Anomalies stop being spooky; they become outliers in a distribution.

The historical challenge has never been a lack of weird reports. It has been the absence of a scalable, scientific filter that can separate credible signals from camera glitches, lens flares, and misidentified aircraft. Human pattern recognition does not scale to millions of objects per year.

So the central question shifts from “Are UFOs real?” to something more technical: how do you search for an object when you do not know its shape, size, propulsion, or brightness profile? How do you design an experiment for a category that might not exist yet in any database at all?

Harvard’s UFO-hunting AI starts by doing something profoundly unsexy: mastering the ordinary. Researchers train it not on sci-fi saucers, but on every known thing that routinely clutters our skies, from flocks of geese to SpaceX Starlink trains streaking across twilight.

Instead of asking the model to recognize aliens, they ask it to memorize normalcy. Birds, airplanes, satellites, helicopters, weather balloons, drones—if humans have flown it or nature has thrown it, the system learns its visual and motion fingerprint.

That demands a “rich dataset,” as Avi Loeb describes it. The Galileo Project’s observatories feed the AI real-world footage of these objects under wildly different conditions, because a 737 at noon and a 737 at dusk look nothing alike to a neural network.

Engineers vary parameters systematically, turning the dataset into a multi-dimensional map of the sky’s behavior. Each labeled example comes with context: time of day, illumination level, angle relative to the sun, atmospheric clarity, and background clutter like clouds or city lights.

So a single satellite pass can spawn dozens of training samples: - Different sun angles glinting off solar panels - Multiple exposure settings from the cameras - Varying degrees of motion blur and noise

Over millions of such instances per year, the AI internalizes the “rules” of our sky. Commercial jets follow predictable corridors and speeds, satellites trace smooth arcs, birds flap and bank in characteristic patterns, and helicopters hover and pivot in ways planes never do.

This baseline of normalcy becomes the project’s real weapon. By saturating the model with what is definitively not a UFO, researchers massively shrink the search space for anything that might be.

Crucially, the system does not become an expert in aliens; it becomes an expert in everything else. Once the Machine Learning model can say with high confidence, “That’s a bird, that’s a plane, that’s Starlink,” what remains are the stubborn residues—the outliers that obey none of the learned patterns and demand a closer look.

Outlier detection sounds abstract, but in Harvard’s UFO hunt it has a concrete job: patrol the tail of the distribution. After the AI memorizes what “normal” looks like—birds, jets, satellites, drones—it treats every new frame as a statistical question: does this belong to the familiar crowd, or does it live far out in the weird, low-probability fringe?

That is exactly what Avi Loeb means when he says, “Then the only question that remains is are there any outliers?” Once the system digests millions of examples of known sky traffic under different lighting, angles, and weather, the UFO problem collapses into a single query: which detections refuse to fit any trained category?

Targets here are not cinematic flying saucers but violations of physics expectations. The system flags objects that exhibit: - Unusual speed relative to altitude and size - Apparent “impossible” acceleration between frames - Motion profiles that ignore known aerodynamic or orbital constraints

Instead of asking “does that look like a craft?” the pipeline asks “does that trajectory match any legal move in the playbook of known objects?” An airplane has a constrained performance envelope; a satellite tracks a predictable orbit; a bird flaps with characteristic frequencies. Anything that jumps between those regimes, or exceeds plausible g-forces, lights up as an anomaly.

This flips UFO hunting from subjective interpretation to objective data analysis. Human observers bring bias, expectation, and pattern-matching errors; a trained model brings probability scores and error bars. If an object’s behavior sits many standard deviations away from the learned norm, it becomes a candidate for deeper scrutiny, regardless of how mundane or exotic it appears.

Projects like the Galileo Project lean on this strategy because their observatories will generate data on millions of objects per year. They need automated triage that elevates only the strangest 0.01 percent. Related efforts in astronomy, such as COSMICA: A Novel Dataset for Astronomical Object Detection, show how curated datasets and outlier-focused models can surface rare, potentially paradigm-shifting phenomena hiding in plain sight.

Harvard astrophysicist Avi Loeb has decided that UFOs are too important to leave to grainy YouTube clips and government leaks. He wants them dragged into the same brutal daylight that exoplanets, black holes, and fast radio bursts now live in: systematic surveys, reproducible data, and code that anyone can inspect.

Loeb’s Galileo Project takes its name seriously. Just as Galileo pointed a new instrument at the sky and rewrote the rules, Loeb is wiring up custom observatories to watch the heavens continuously, then feeding that firehose into Machine Learning pipelines. One array sits in Massachusetts, another in Pennsylvania, and a new site in Las Vegas is coming online, each expected to log data on millions of objects per year.

Every station stacks multiple sensors on a single patch of sky: optical cameras, infrared, radar feeds when available, and precision GPS timestamps. That fusion matters, because a weird shape in one band often turns into a mundane satellite once you track its motion, spectrum, and altitude together. Only objects that stay weird across modalities survive as genuine anomalies.

Core to the mission is a hard line on transparency. Loeb’s team commits to publishing all raw and processed data, along with detection algorithms, for full peer review. No classified feeds, no special-access reports, no “trust us, we saw something”: if the system flags an object, other scientists get to poke at the pixels and the code.

That stance directly challenges decades of UFO taboo inside astronomy. Instead of reputational risk for even mentioning “unidentified,” Galileo reframes the whole topic as a standard outlier-detection problem: same math used in fraud detection or cybersecurity, now aimed at the sky. Anomalous trajectories, accelerations, or light curves become just another dataset, not a career hazard.

By committing to open methods and reproducible statistics, Loeb wants to shift the conversation from late-night speculation to verifiable science. Either the anomalies collapse into better-understood drones, balloons, and satellites, or a residue of truly unexplained events survives repeated scrutiny. For the Galileo Project, both outcomes count as progress.

Concrete hardware now backs Avi Loeb’s big idea. The Galileo Project is quietly building a distributed network of AI-assisted observatories, starting with full instrument stacks in Massachusetts and Pennsylvania and a newly constructed site on the outskirts of Las Vegas, Nevada. Each station operates as a self-contained data factory, designed to watch the sky continuously and feed algorithms a firehose of labeled reality.

Instead of a single flagship telescope, Galileo favors modular rigs. Arrays of high-resolution optical cameras track visible light, while paired infrared sensors watch heat signatures that conventional astronomy gear often ignores. Radar receivers and radio antennas can slot in to catch transponders and communication signals, helping separate stealth jets from weather balloons from something stranger.

Every site aims to log millions of sky objects per year, not as blurry anecdotes but as structured, queryable data. That means recording not just images, but time-stamped trajectories, apparent velocities, accelerations, and spectral fingerprints. Loeb’s team wants each frame to become a training example: one more entry in the world’s most exhaustive catalog of “normal.”

To keep up, the observatories push intelligence to the edge. Racks of on-site GPU and TPU hardware run Machine Learning models directly beside the cameras, performing real-time object detection and classification. Instead of streaming raw 4K footage 24/7, the systems compress reality down to metadata, thumbnails, and anomaly scores.

A typical Galileo stack could include:

• 1High-frame-rate 4K or higher optical cameras with wide and narrow fields of view
• 2Short-wave and mid-wave infrared sensors for thermal profiles
• 3All-sky cameras for continuous horizon-to-horizon coverage
• 4Precision GPS and inertial sensors for exact pointing and timing
• 5Local AI servers with multi-GPU nodes for inference and retraining

Las Vegas offers a particularly revealing testbed. Clear desert skies, frequent air traffic, drones, and bright city light pollution create a noisy environment that stress-tests the models. If the software can reliably distinguish helicopters, private jets, fireworks, and SpaceX Starlink trains over Nevada, it stands a better chance of spotting a legitimate outlier anywhere else.

Scaling this network turns Earth into a kind of planet-sized sensor array. Each additional node increases sky coverage, weather diversity, and viewing angles on the same event. Stitch that together, and Galileo stops being just a UFO camera project and starts looking like a real-time, AI-driven census of everything that moves above us.

Rocks dominate telescope time. Ask an astronomer what a weird dot in the sky might be, and the reflex answer still leans on “icy rocks” and dusty comets, not hardware. Avi Loeb argues that this bias hardwires a blind spot into both human judgment and the algorithms that inherit our training labels.

Traditional surveys optimize for natural categories: asteroids, comets, meteors, satellites. Loeb wants Machine Learning models that also treat “technological objects” as a first-class hypothesis, not an afterthought. That means encoding how engines, panels, or structured alloys might scatter light, heat up, or maneuver in ways no rock can.

Because we do not have alien artifacts in a lab, Galileo researchers lean on synthetic data. They can simulate specular glints from flat surfaces, non-gravitational accelerations from controlled thrust, or flickering light curves from rotating trusses. Those patterns then join training sets alongside real footage of birds, planes, helicopters, and SpaceX hardware.

ʻOumuamua turned this debate radioactive. Detected in 2017, the interstellar visitor showed no visible coma, had a highly elongated shape estimate, and exhibited a small but persistent non-gravitational acceleration. Comet specialists defaulted to exotic outgassing; Loeb asked whether a thin, possibly artificial object—solar sail, shard, or debris—might better fit the oddities.

That fight exposed how prior beliefs steer classification. If you assume every unknown is natural, you will stretch comet physics until it squeaks. If your AI only trains on rocks and ice, it will contort anomalies back into those buckets, because the model literally does not know “technology” exists.

Loeb’s mantra: you can’t judge by its cover, whether you are a grad student or a convolutional neural network. Galileo’s pipeline focuses less on shapes and more on performance envelopes—speed, acceleration, trajectory kinks, and energy budgets. An object that appears mundane but pulls 100 g turns or climbs against drag without visible propulsion lights up the anomaly score.

Researchers in adjacent fields already benchmark systems on unusual optics, as in Benchmarking Deep Learning-Based Object Detection Models on MobilTelesco. Galileo aims for a similar rigor: expose models to both ordinary clutter and hypothetical machines, then demand they flag whatever does not behave like a rock or a rocket.

Human skywatchers come with baggage. We get tired, bored, and distracted. Our brains hallucinate patterns—pareidolia turns Venus into a UFO, a flock of geese into a flying triangle, a lens flare into an alien craft.

AI does not blink. Once Avi Loeb’s Galileo Project spins up, its cameras and sensors stream data 24/7, in visible, infrared, and sometimes radio bands, logging millions of objects per year from sites in Massachusetts, Pennsylvania, and Las Vegas. No coffee breaks, no night shifts, no “I thought it was a plane.”

Instead of asking humans to decide what looks weird, the team teaches Machine Learning systems what looks boring. They feed in huge datasets of birds, airplanes, helicopters, drones, satellites, and weather phenomena, captured at different times of day, illumination levels, and angles relative to the Sun. The model internalizes a high‑dimensional definition of normal sky behavior.

Bias usually creeps in when observers expect a certain kind of UFO—saucer, tic‑tac, glowing orb—and unconsciously filter everything else out. The Galileo AI has no such mythology. It does not know what a “proper” alien craft should look like, or that comet experts prefer “icy rocks” over technological debris.

Every new event gets reduced to a data signature: shape, spectrum, trajectory, speed, acceleration, flicker pattern, and context. The system compares that signature against its learned baseline of normalcy and assigns a probability that the object belongs to a known class. Low probability means “outlier,” not “alien,” but crucially, it means “do not ignore.”

This pipeline flips the usual script. Humans no longer cherry‑pick anecdotes from blurry videos. Algorithms promote candidates purely because the numbers say they do not fit—forcing scientists to confront whatever the sensors actually saw, rather than what they expected to see.

ʻOumuamua and Borisov did not announce themselves to us; astronomers tripped over them. Surveys like Pan-STARRS and amateur networks caught these interstellar visitors because they happened to be looking at the right patch of sky at the right time, not because anyone ran a dedicated interstellar search. Loeb argues that this “cosmic luck” model all but guarantees we miss most of what flies through the solar system.

Current discovery pipelines still behave like security cameras: always on, mostly passive, and tuned for routine background motion. They excel at finding near-Earth asteroids and comets, not objects with bizarre trajectories, non-gravitational accelerations, or strange reflectivity profiles. Anomalies like ʻOumuamua’s odd light curve and unexplained push from sunlight slip through as statistical afterthoughts.

Loeb’s pitch is blunt: swap serendipity for systematic hunting. Train Machine Learning models on millions of “normal” objects—asteroids, comets, satellites, space junk, aircraft—then deploy them as real-time filters on wide-field telescopes and all-sky cameras. Anything that falls outside the learned performance envelope gets flagged instantly for follow-up.

He claims that approach could boost the discovery rate of interstellar objects by “two orders of magnitude.” Instead of one or two confirmed interstellar visitors per decade, AI-assisted surveys could log dozens or even hundreds each year. That jump comes not from bigger mirrors, but from smarter triage on the data torrent we already collect.

This flips astronomy’s stance from reactive to aggressively proactive. Rather than waiting for something weird to wander across a CCD, algorithms can predict likely inbound trajectories, cross-match tracks across observatories, and prioritize scarce pointing time. Telescopes stop behaving like static webcams and start acting like a coordinated, predictive sensor network.

Loeb frames this as the opening phase of an “interstellar gold rush” for data. Early adopters with AI-augmented observatories will stake the first claims on high-value anomalies: objects with unusual spectra, strange accelerations, or orbits that scream “not from here.” In a sky suddenly mined for outliers, every odd blip becomes a potential strike.

AI hunting for UFOs sounds like science fiction, but the Galileo Project sits inside a much larger revolution. Across labs and observatories, Machine Learning is turning pattern recognition from a human bottleneck into an industrial process, compressing years of analysis into minutes and surfacing structures no one knew to look for.

DeepMind’s AlphaFold did for biology what Avi Loeb wants AI to do for the sky. By predicting the 3D shapes of roughly 200,000 proteins at launch—and later scaling to hundreds of millions—AlphaFold jump-started drug discovery, enzyme design, and basic cell biology in a way brute-force computation never could.

Similar step changes are rippling through astrophysics. Projects cataloging billions of stars and galaxies—Gaia, Pan-STARRS, and soon the Vera C. Rubin Observatory—generate petabytes of data that no human team can fully inspect. AI models now classify supernovae, map dark matter via gravitational lensing, and flag transients that exist for hours before fading forever.

Galileo’s UFO-hunting stack plugs directly into that wave. Training on birds, airplanes, satellites, helicopters, and drones across illumination conditions turns “normal” sky clutter into a compressed model, so anything off-distribution—bizarre acceleration, odd spectra, impossible trajectories—pops out as a statistical scream rather than a blurry maybe.

Crucially, this is a step function, not a linear upgrade. Instead of adding more grad students to sift telescope images, researchers deploy models that learn from millions of labeled examples and then generalize. The shift mirrors AlphaFold: once the system understands the space of possibilities, novel cases become cheap to spot.

Harvard’s own astronomers lean into this transition. The Center for Astrophysics runs dedicated efforts in AI-driven classification, anomaly detection, and simulation; its overview at Machine Learning | Center for Astrophysics | Harvard & Smithsonian reads less like a niche program and more like a new operating system for the field.

Wes Roth and Dylan Curious frame Galileo as one node in a broader tech map. Their “AI POD,” now past 190 episodes, jumps from AlphaFold to massive tech ontologies like Cosmos 1.0’s 23,000 technologies, sketching a future where AI does not just analyze data, but actively shapes which experiments, telescopes, and even spacecraft humanity builds next.

Someday a Galileo Project camera will flag an object that refuses to fit. The Machine Learning system will compare it against millions of labeled examples—birds, drones, jets, Starlink trains—and fail. In that moment, the outlier detection pipeline stops being an academic exercise and becomes a discovery channel.

Because Loeb’s team commits to open, peer-reviewed data, that alert will not disappear into a classified inbox. Raw frames, tracking metadata, and model outputs will go to public archives and journals, not a black-budget program. Astronomers, atmospheric scientists, and defense analysts worldwide will get the same pixels.

If independent teams confirm a non-human, non-natural origin—say, a craft with accelerations far beyond known propulsion and no plausible sensor artifact—that result detonates across science. Planetary science, plasma physics, propulsion engineering, and information theory will all need new branches to explain a working technology stack that did not originate on Earth. Funding priorities at agencies like NASA, ESA, and NSF will pivot overnight toward “contact-ready” research.

Society will not process that shift cleanly. Religious traditions, political narratives, and national security doctrines all assume humans occupy the top technological rung. A verified non-human artifact, backed by open data and reproducible analysis, undercuts that assumption more forcefully than any speculative UFO story.

Harvard’s network—Massachusetts, Pennsylvania, Las Vegas—exists to make that moment survivable by science. Every frame already carries calibration data, timestamps, and cross-sensor verification so that, if something impossible appears, skeptics can attack the pipeline, not the witnesses. Peer review becomes the first contact protocol.

Once a single object graduates from “unknown” to “confirmed non-human technology,” the training set changes forever. Known objects will now include at least one entry tagged “technological, non-terrestrial.” From that point on, every new model will learn not just what humans build, but what someone else already did.

How does AI detect unusual sky objects or UFOs?

Instead of looking for UFOs directly, the AI is trained on a massive dataset of known objects like birds, airplanes, and satellites under all conditions. It then flags anything that doesn't fit this known behavior, identifying 'outliers' with unusual characteristics.

What is the Galileo Project?

Led by Harvard astrophysicist Avi Loeb, the Galileo Project is a scientific initiative to systematically search for evidence of extraterrestrial technological artifacts. It uses a network of AI-powered observatories to analyze the sky and commits to publishing all findings for peer review.

Why is this AI approach better than traditional methods?

Traditional searches are often sporadic or rely on chance discoveries. This AI method provides tireless, 24/7 monitoring and removes human bias, allowing for a systematic and objective hunt for anomalies at a scale that could increase the discovery rate by 100 times.

What kind of data is used to train the AI?

The AI is trained on a rich dataset of real-world observations of all familiar natural and human-made objects. This includes birds, airplanes, helicopters, and satellites, captured at different times of day, with varying illumination levels, and from multiple angles.