Partners

Our partnership with Frothly

Since 2021 we have worked with Frothly, the independent institute in San Francisco that runs the Turing Registry, on a question our founder’s company raised and cannot answer by itself: who should check the machines inside people’s eyes?

Partners since
27 September 2021
Frothly programmes
Blackwall (evaluation), Lazarus (preservation), Synthetic Media (research)
Our grants to Frothly
4 grants since 2021, ¥379 million in all
Conditions on findings
None. Frothly accepts no funding that comes with conditions on what it finds, and we offer none.
Contact
Dr Tachibana Shion, Director of Research and Standards, Chiba
Why Frothly

An optical implant is a computer that sits between a person and everything they see. Somebody independent ought to be able to look inside it.

Kiroshi Opticals, the company our founder started as a lens workshop in 1952, now makes cybernetic eyes and retinal implants used by about 3.4 million people. The software in them decides what to sharpen, what to label and what to leave out. For people who were blind, that software is the difference between a doorway and a wall. For everyone who wears one, it is a set of choices made on their behalf by a company in which the Foundation is the largest shareholder.

We could not credibly check that software ourselves, and we did not want to pretend we could. Frothly has spent four decades studying what happens to public life when machines decide on our behalf. It holds no shares, sells no products and publishes what it finds. In 2021 we asked whether it would work with us, on its terms.

Evaluation

Kiroshi Opticals and Blackwall

Blackwall is Frothly’s programme of independent evaluations of AI systems before they are deployed. Laboratories submit voluntarily, and every evaluation ends in a public report, including findings the laboratory would rather were not published.

In 2023, as Kiroshi Opticals’ largest shareholder, we asked the company to submit its perception models to Blackwall before release. It agreed on 30 October 2023. Since then it has submitted two: SHIKAI 3, the model in implants sold from 2024, and SHIKAI 4, from 2026. That participation, together with this research partnership, is why Kiroshi appears among the institutions listed on Frothly’s home page.

We pay for none of Blackwall’s evaluations of Kiroshi Opticals systems, and we do not see the reports before they are published. When Blackwall’s evaluation of SHIKAI 4 found that the model was markedly less reliable at recognising white canes and guide dog harnesses in low light, we learned of it on the day of publication, along with everyone else. Our statement set out what we asked the company to do next.

Joint research

What we fund at Frothly

Four grants, each to a named Frothly programme for a named piece of work, each listed in our grants database.

Foundation grants to Frothly, 2021 to 2026
YearsProjectFrothly programmeGrantStatus
2021–2023Who checks the eye? A scoping study of independent oversight for optical implantsFrothlyKF-2021-0317
¥38 million, 27 September 2021
Completed. Report published 2023.
2023–2026Open methods for evaluating perception models in medical implants, with a public low-vision test setBlackwallKF-2023-0241
¥96 million, 14 September 2023
Test set KF-LV 1.0 released March 2025. Methods paper in review.
2024–2027Preserving and documenting the firmware of discontinued optical implantsLazarusKF-2024-0388
¥180 million, 9 December 2024
SEIRAN 2 and Raven Corvid firmware deposited. Catalogue records follow the archive’s re-cataloguing.
2025–2027Seeing what is not there: how overlays in ocular displays change what wearers believeSynthetic Media, ChibaKF-2025-0102
¥64.5 million, 30 June 2025
Fieldwork under way in Chiba and Osaka, led by Dr Peter Riviera.

The Lazarus archive is re-cataloguing its records after a migration of its index in August 2026, so records for the deposited firmware will appear in its online catalogue as that work reaches them.

Reading

Frothly research we rely on

  • Through a Glass, Darkly

    Frothly paper, May 2026

    Found that the explanations recognition systems give for their decisions were faithful no more than about half the time. KF-OS 2 now requires that any explanation an implant shows its wearer has passed intervention testing.

  • On Artificial Owls

    Frothly report, September 2026

    Recommends that device makers sign images at the moment of capture, in ways that protect the identity of the person capturing them. An optical implant is the most personal capture device there is; KF-OS 3 is our attempt to do this properly.

  • Half-Life: What Happens to Our Data When We Die

    Frothly policy brief, August 2026

    Our guidance on what should happen to the stored data in an implant when its wearer dies follows the brief’s framework. Lazarus’s digital estates pages show it in practice.

  • Ethical Constraints Are Load-Bearing

    Frothly essay, June 2026

    Required reading for anyone at the Foundation who works on implant firmware. The working group reviewing KF-OS 1 this year has asked for it to be quoted in the revised standard’s foreword.

Frothly’s research library

The Registry

The Turing Registry question

The Turing Registry, which Frothly runs from Zürich, is a public record of autonomous systems: who built each one, what it is licensed to do and who answers when something goes wrong. People often ask why Kiroshi Opticals has no filings on it.

The company’s position is that the perception model in its implants is a component of a medical device, which informs its wearer but does not act in the world, and that medical device regulators in the twenty-three countries where it sells implants are the right authorities to oversee it. That position is not unreasonable, and it is not ours to decide.

Our position is that the question deserves a public answer rather than a default. We have asked the company, in writing, to keep the matter under review as its implants do more, and we have published the request. When we wrote KF-OS 2, our standard on bystander privacy, we cited the Registry’s 2023 revocations of TR-08333 and TR-08332, two systems licensed to keep or extract people’s engrams, which lost their licences for doing so without consent. They are the reason the standard forbids an implant to retain images of the people its wearer looks at.

Limits

What the partnership is not

  • We hold no seat on Frothly’s board, and no part in its Blackwall panels or Registry decisions.
  • We are not told in advance what Blackwall will publish about Kiroshi Opticals, and we have never asked to be.
  • Frothly’s Scramble programme evaluated six commercial recognition systems for Through a Glass, Darkly on condition of anonymity. We do not know whether a Kiroshi Opticals system was among them, and we have not asked.
  • Frothly staff who work on our grants are paid by Frothly, not by us.

Frothly is at frothly.org. Its programmes are described on its programmes page, and its next Citadel conference meets in Vancouver in April 2027 (details).