'Our Laptops Aren't Racist,' HP Says

Updated slfisher 1 395 Views

Hewlett-Packard is scrambling to respond to an issue brought up by a Dec. 10 YouTube video demonstrating that the facial tracking software used in some of its laptops doesn't recognize black people.

HP responded to the issue yesterday in its blog, and it's starting to go viral today.

In the light-hearted but pointed video, black "Desi" and white "Wanda" show how the software tracks Wanda but not Desi, even though they're in the same room, at the same angle, with the same lighting. The tracking works when Wanda enters the frame, and stops working when Desi enters.

A mortified HP blamed the problem on insufficient contrast between the eyes and the skin of the upper cheek and nose. "We believe that the camera might have difficulty “seeing” contrast in conditions where there is insufficient foreground lighting," HP said, referring people to about optimum lighting for facial-tracking software.

Meanwhile, the company is pledging to work on the problem with its partners.

Dani AI

Generated

This thread highlights a technical failure mode that shows up again and again in face-detection / tracking systems: uneven performance across skin tones and demographic groups. Landmark academic work measured large accuracy gaps by skin tone and gender, and NIST’s large-scale FRVT testing documents demographic differentials across many commercial algorithms. These are engineering problems with real social consequences, not just PR noise. Gender Shades (MIT) — NIST FRVT: Demographic Effects. (media.mit.edu)

Two failure sources commonly interact. First, image formation: consumer webcams and auto-exposure systems can underexpose darker faces because metering and sensor pipelines historically bias midtones; this is a long-standing photography issue (see Kodak “Shirley” calibration history). Second, model training: detectors and trackers trained on datasets skewed toward lighter-skinned subjects will generalize poorly to underrepresented groups. Both hardware and data matter. . (kawc.org)

Practical engineering guidance (ties to ’s testing idea): create a repeatable “clinical” validation matrix (Fitzpatrick I–VI, sexes, age ranges, glasses/occlusion, multiple poses, front/side/back lighting, exposure variants). Track subgroup metrics: detection recall, false positives, landmark error, and tracking continuity. Use balanced benchmarks (Pilot Parliaments / Gender Shades) and fairness toolkits to quantify gaps, and mitigation tools like IBM’s AIF360 and Microsoft’s Fairlearn to evaluate and reduce disparities. Mitigations: add controlled illumination (or IR for eye-tracking where available), adaptive preprocessing (CLAHE/gamma), targeted retraining with diverse samples and synthetic lighting augmentation, and clear diagnostics/updates. Pilot Parliaments / Gender Shades resources — AI Fairness 360 (IBM) — Fairlearn. (montrealethics.ai)

On the public-facing side (ties to ): don’t treat isolated viral clips as only PR — they’re reproducible test cases. For end users, simple fixes often help: face the light source, avoid strong backlighting, increase webcam exposure/exposure compensation where possible, and install driver/firmware updates; for vendors, publish test procedures and ship fixes. Exposure / exposure-compensation primer. (nist.gov)

Member Avatar for MktgRob
MktgRob 7 Nearly a Posting Virtuoso

Did they try any other people with unique features based on ethnic backgrounds? If what HP says is true then I would imagine there would be difficulties with people from around the Pacific Rim and Asia as well. I am not saying who is right or wrong here but it is amazing how one little video on youtube can be used to put a label on a corporation or group with no other imperical evidence presented.

Member Avatar for InsightsDigital
InsightsDigital 57 Posting Virtuoso

Perhaps HP and other technology companies need to pay heed to what pharmaceutical companies have to do while performing necessary clinical trial runs - pharma companies have to test the product with different groups including control to make sure the drug works. There are usually 3 rounds before the drug gets FDA approval and many times the drug might demonstrate protective or damaging effects on certain racial segments. The lesson here is that technology companies like HP should test its products with different subjects and debugging it before it officially releases the product - unless the product is in beta stage.

Be a part of the DaniWeb community

We're a friendly, industry-focused community of developers, IT pros, digital marketers, and technology enthusiasts meeting, networking, learning, and sharing knowledge.

Edit Preview