Methodology
One evidence view, three layers, clear limits
Viewers watch your creative. Webcam AI estimates their emotional responses, expression and posture at each moment; their comments explain why. Estimates guide improvements — they are not certainties.
Evidence layers
Layer 1
Emotion, expression and posture estimates
With consent, webcam AI estimates emotional responses such as positive response, interest or confusion from facial expression, and notes posture cues such as leaning in or back, aligned to the moment in the creative.
Layer 2
Viewer feedback and actions
What people say and do — comments, answers, pauses, rewinds. Testimony explains why an estimate moved.
Layer 3
Optional webcam-derived signal
With consent, a heart-rate estimate derived from subtle colour changes in facial video (remote photoplethysmography, rPPG). It is aligned to task events so researchers can see where something changed.
Limitations
What estimates can’t tell you
- A pulse estimate cannot identify what someone feels or why. Similar changes can come from effort, surprise, interest, posture or simply talking.
- Signal quality depends on lighting, camera, movement, skin tone coverage, framing and connection. Segments that don’t meet a quality threshold should be flagged or excluded.
- Webcam estimates are not equivalent to medical or wearable devices and are not used for diagnosis or any clinical purpose.
- Small studies show directions, not certainties. Results should be confirmed with adequate samples before major decisions.
- Emotion values are estimates, not certain knowledge of what anyone felt. No detection of stress, intent or truthfulness is claimed, and estimates are checked against what viewers did and said.
Research evidence vs. product validation
An independent scientific review found that facial movements alone are not a reliable basis for inferring how a person feels across contexts and cultures: Barrett et al., 2019, Psychological Science in the Public Interest (PubMed 31313636). That review is general research evidence about emotion inference. It is not a validation study of Optimizing AI, and no independent validation of this product is claimed here.
Draft
Research use boundaries
Draft boundaries for how studies are intended to be run. They describe intent, not a legal or compliance commitment, and are confirmed per study.
- Designed for voluntary product and creative research. Optimizing AI is not offered for monitoring employees or students or making decisions about individuals.
- Participation is voluntary. The webcam signal is opt-in, and participants can take part without it.
- Studies test a creative, concept, video or flow, optionally against a revised version. Results describe how groups responded to a design, not a judgement of any one person.
- Estimates are reviewed alongside viewer feedback and actions, never on their own.
Participant disclosure for the samples on this site
This sample uses fictional data drawn in your browser. It does not access your camera or microphone, and the interactive controls on this page make no network requests.
Data handling
Questions to settle before a pilot
Data handling is agreed per study. These are the questions we suggest every team asks — of us and of any research tool. Specific answers for your study are confirmed in writing during setup, not assumed from this page.
- Q1Is the webcam signal optional for every participant, and can they take part without it?
- Q2How and when is consent requested, and can a participant withdraw during or after a session?
- Q3Is raw video stored, or only derived values? For how long, and where?
- Q4Who on your team and ours can access session data, and how is access logged?
- Q5How are low-quality segments identified and shown in results?
- Q6How is participant data deleted at the end of a study, and how is that confirmed?
- Q7Which agreements (for example a data processing agreement) are needed for your organisation?
Further reading
The founder’s public notes on the approach, evidence and scope are at hannesbend.com/ventures/optimizing-ai. To see how the layers look together, open the sample study.
A 30-minute conversation, booked through Calendly.