Why AI Will Never Be Perfect: Lex Fridman's Case for Keeping Humans in the Loop

Fridman opens with a prediction: the learning-based approaches that have succeeded over two decades will keep improving and dominate real-world applications, displacing hand-tuned optimization methods. If that is true, smarter AI comes from improving two things, machine learning (optimizing network parameters from data) and machine teaching (optimizing which data the algorithms learn from).

He compares machine teaching to the way parents and teachers give a child sparse but well-chosen information. Selecting the data to learn from is, in his view, a critical research direction for building systems that work in the real world. He also expects a pure learning approach to hit a wall unless humans are integrated deeply into the process.

The core reason is generalization. A learning system must generalize across all of reality from a very small subset of samples, leaving permanent uncertainty and incomplete information. As a result, such systems will not be provably safe unless extremely constrained, will not be provably fair, and will not be perfectly explainable to human supervisors, so supervision must be constant.

Fridman describes conventional deep learning as human-out-of-the-loop: people annotate a dataset, the system trains on it, and in operation the human is only the recipient of the service. The human-centered version integrates people in both phases, drawing on the wisdom of individual experts and of crowds.

He distinguishes objective annotation, where one person can label a sample as in ImageNet, from subjective annotation, where crowds converge on fuzzy emotional or ethical questions nobody can answer alone. In operation, a single driver might supervise an autonomous vehicle, while ethical decisions in areas like medical diagnosis need crowd intelligence.

On the training side, he proposes reversing the usual flow. Rather than humans doing brute-force annotation of datasets such as ImageNet, Cityscapes or YouTube action data, the model queries humans with questions, a field called machine teaching. The challenge is to cut the annotation needed by several orders of magnitude.

In operation, the system must report its uncertainty, and when it falls below a threshold, as with a costly financial decision or a decision affecting human life, human supervision is sought. Fridman believes this will be the defining mode of operation for AI systems in the 21st century.

Fridman lists five grand-challenge areas: machine teaching, reward engineering, human sensing, interaction experience, and safety and ethics. For machine teaching the goal is to learn from very few examples, the way students need only a few critical ones. Near-term tools include active learning, data augmentation, one-shot, zero-shot and transfer learning, and self-play.

He proposes concrete grand challenges: training only on Wikipedia text and images yet achieving state-of-the-art object detection on COCO, or reaching 0.3% error on MNIST from a single example per digit. Humans can do the latter after studying one example of each character in a new script.

The second arm is reward engineering, injecting humans into the definition of good and bad during learning. He shows an OpenAI example in which a boat-racing RL agent discovers that collecting green turbos earns more reward than finishing the race. A human supervisor would re-engineer the reward so the agent finishes, and continuous monitoring of this kind is a near-term direction.

His provocative grand challenge for the ambiguous gray areas of ethics is replacing the US Congress with a human-centered AI system, which he compares to a recommender system for the people's beliefs. In Q&A he clarifies that it should automate the system as it is supposed to work, with more transparency, and with constant human supervision rather than letting AI run the country.

Human sensing is the perception problem of turning video, audio and text into the physical, mental and social state of a person, where deep learning shines. Fridman argues that emotion recognition is still unsolved: current systems detect caricatured expressions such as smiling or frowning rather than real emotion.

He offers a clean test: a binary classifier with 95% accuracy on whether a person wants to be left alone, trained on 30 days of data. We are very far from that, especially in integrating information over long periods of time.

On interaction, he compares semi-autonomous driving systems: Volvo's small icon, Tesla's large display and Cadillac Super Cruise's eye-tracking camera. He cites over a billion Tesla miles studied at MIT and 10-plus million autonomous Waymo miles, and says the next challenge is hundreds of billions of semi-autonomous miles with teenagers and older drivers. The Alexa Prize socialbot challenge is, for him, the Turing test reimagined.

On safety, he stresses supervising life-critical and ethical decisions such as loans or criminal penalties, avoiding bias in perception systems like face recognition, and controlling outliers as well as the mean. Everything boils down to calibrated uncertainty: a model trained on cats and dogs shown a giraffe should say it doesn't know, whereas today it would confidently say dog.

Fridman says each perception topic in his full slides follows the same template: motivation, utopia and dystopia, then five papers from seminal work to future directions, and open problems. For faces, he separates detection, recognition (which of billions of identities is this) and verification (is it Lex or not, as when unlocking a phone).

Faces are hard because of lighting and pose, but also because identities can look very similar and most people have few samples. Systems must be invariant to hairstyles, weight changes, beards, glasses and makeup. The stakes of accuracy are high for security.

The utopian vision is using your face as a passport and replacing passwords and cards, with "Face Pay" instead of Apple Pay. The dystopian version is complete loss of privacy, impossible to hide from society. The middle path is simply a convenient way to unlock a phone.

Technically, DeepFace brought deep networks to near-human performance on Labeled Faces in the Wild, and FaceNet learned an embedding where Euclidean distance reflects identity, which is essential when each person has few examples. Open problems include recognition beyond frontal bounding boxes using context, interpretability and bias of black-box systems, and privacy.

Activity recognition classifies what people are doing from images or video, giving context for understanding them. Fridman notes that it is hard because video is huge, motion is difficult to quantify from a 2D view, and activities are ambiguous: when does one start and end? The utopian view is richer human-robot interaction, and the dystopian one is robots that sever the relationship.

He traces progress through C3D 3D convolutions and two-stream networks that combine RGB with optical flow, plus LSTM-based models. Future directions include predicting the future from images to augment data. Open problems include localizing activities in long videos and handling context such as other objects and people in the scene.

Body pose estimation localizes the joints of the skeleton. It matters for reading body language and for robots handing someone a beer or cars checking whether a driver's hands and head position allow them to take control. The dark flip side is that localizing joints could help robots hurt people more effectively.

From DeepPose in 2014, which removed hand-crafted features, he moves to CMU's real-time multi-person approach that detects all joints first and then stitches them together with parts affinity fields. His open problem is that pose is part of movement through time, and temporal aspects of behavior, from silly walks to emotion, are mostly ignored.

Fridman contrasts black-box systems such as Tesla Autopilot, which show no uncertainty and ask us to have faith, with his group's arguing machines idea. An ensemble of independent systems is used, and when they disagree, human supervision is sought. He links it to OpenAI's work on safety through machine debate, which avoids giving ultimate power to any one decision maker.

In an instrumented Tesla, a second end-to-end vision system makes steering decisions, and disagreement between the two predicts with high accuracy the moments when drivers chose to disengage because they were uncomfortable. This flags risky, challenging situations.

The same idea applies to vision. ResNet (8% error) and VGG-16 (10%), trained on ImageNet, can argue, and when disagreements go to a human the error drops to 2.8%. In examples, a wine bottle is called a paper towel and a seat belt, and a mailbox becomes a traffic light and a garbage truck. Fridman notes this would matter if the error meant loss of life.

He closes the demonstration with MIT's human-centered autonomous vehicle, which passes control back and forth based on the driver's activity. His point is that the experience should be fun and enriching, and that many problems in this space are untouched.

Fridman says most of the robotics and machine learning communities try to make AI perfect before releasing it to the world. He expects AI systems to be imperfect for the next hundred years, so the task is to optimize humans and AI working together, both flawed but enriching each other.

He uses an analogy of parasitism versus symbiosis. Today's training injects costly, offline brute-force human labor, whereas the future he wants has models improving as a side effect of interacting with humans. When annotation happens naturally through interaction, learning can scale to the degree real-world problems demand.

Achieving this spans neuroscience, cognitive science, psychology, sociology, game theory, human factors and design, natural language processing, HCI, deep and reinforcement learning, and robotics hardware. All have to work together to integrate the human during both learning and operation.