Industries / Retail, hospitality and personal services / Funeral services
Memorials
Memorials is how the dead are named and remembered after the funeral: headstones and plaques, and the online records, pages and recordings that keep a life in view.
5
Milestones achieved, from 1918 to 2024.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1918
One headstone for all
On a 1918 report by Sir Frederic Kenyon, Director of the British Museum, the Imperial War Graves Commission, now the Commonwealth War Graves Commission, decided that war graves would carry headstones of one uniform design whatever the rank or religion of the dead, lettered in a typeface drawn for them by Leslie MacDonald Gill. Three trial cemeteries in France, designed by Reginald Blomfield, were built by 1920.
LessonA special committee set up by the Commission reported in 2021 that outside Europe the rule was not kept: an estimated 45,000 to 54,000 casualties, mostly Indian, East African, West African, Egyptian and Somali, were commemorated unequally, and at least 116,000, possibly 350,000, were not commemorated by name or perhaps at all.
Credit
Sir Frederic Kenyon; the Imperial War Graves Commission
1995
Graves listed online
Jim Tipton started Find a Grave in 1995 as a website about visiting the graves of famous people, then opened it to everyone; volunteers add burial details, photographs, GPS positions and life stories. Ancestry bought it in 2013, and the site says its community has created over 265 million memorials.
Credit
Jim Tipton; Find a Grave's volunteer contributors
2015
An heir for the account
In February 2015 Facebook let people name a legacy contact, who can pin a post to a memorialized profile and, if allowed, download an archive of what the person shared, but cannot log in as them or read their messages; before this, a memorialized account could be viewed but managed by no one. Apple added Legacy Contact in iOS 15.2, released on 13 December 2021: the contact needs an access key and a death certificate, and gets no access to passwords or payment details.
Credit
Facebook (now Meta); Apple
2023
Machines read the stones
BillionGraves, a site of gravestone photographs taken by volunteers, began transcribing names and dates by machine learning in 2023, after a volunteer warned that more than 2 million photographed stones were waiting for someone to type them. In the first weeks the system completed more than half a million transcriptions; the company says they are right about 80% of the time and that volunteers check the rest.
LessonFaded letters drop out (its own example read "Alice" as "Lice"), so a person still has to check each stone.
Credit
BillionGraves
2024
A memorial company fails
StoryFile, founded in Los Angeles in 2017, recorded question-and-answer video interviews so that people could later hold a simulated conversation with the person on screen. On 5 May 2024 it filed for Chapter 11 bankruptcy protection with $1.5 million in assets and $10.5 million in liabilities, its interim chief executive stating that it had "continued to hemorrhage cash".
LessonA recorded memorial lasts only as long as the company and the file format that hold it.
Credit
StoryFile; filing reported by the Westfair Business Journal
Watch for
AI reads the gravestone
Find a Grave began beta-testing Smart Transcribe on a test server in June 2026: users upload gravestone photographs, and AI checks for duplicate records and transcribes the stone for the user to review and edit. MyHeritage launched Scribe AI in March 2026 and says it transcribes gravestone inscriptions and interprets the symbols carved on them.
Achieved whenan outside party publishes error rates for either tool on photographs of worn stones.
Claimed by
Find a Grave (Ancestry), June 2026; MyHeritage, March 2026
A likeness from three minutes
2Wai released an iOS beta on 11 November 2025 of HoloAvatar, which it says builds a conversational video avatar of a person, including a dead relative, from as little as three minutes of footage, audio and text, runs on the phone, and answers only from data the user has approved. Its launch video drew wide criticism.
Achieved whenan outside security or research group confirms that conversations are processed on the device and that the avatar's answers stay within the approved material.
Claimed by
2Wai (co-founders Calum Worthy and Russell Geyser), November 2025
Open
01
A worn-stone reading test
A public test set of photographs of carved inscriptions with exact answer transcriptions, graded from crisp to badly worn, for scoring the AI tools now offered for reading gravestones. To keep real families out of it, the inscriptions are made-up names, dates and verses cut by a monument mason into offcuts of granite, marble, slate and sandstone, then aged with repeated salt and freeze-thaw cycles in a laboratory, the kind of accelerated weathering used to grade building stone, and photographed in flat and low, raking light.
The briefCredit · unclaimed02
Safety test for griefbots
A public suite of scripted test conversations for chatbots and video avatars that imitate a dead person (often called griefbots or deadbots), run only with made-up people or consenting living volunteers, that checks the failings researchers have warned about: whether the bot says it is an AI when asked, whether it promotes products in the voice of the dead, whether it stops and can be retired when a user asks, whether it accepts a child as a user, and whether it can be built from someone's data without their consent.
The briefCredit · unclaimed
Open 01 · the brief
A worn-stone reading test
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test set of photographs of carved inscriptions with exact answer transcriptions, graded from crisp to badly worn, for scoring the AI tools now offered for reading gravestones. To keep real families out of it, the inscriptions are made-up names, dates and verses cut by a monument mason into offcuts of granite, marble, slate and sandstone, then aged with repeated salt and freeze-thaw cycles in a laboratory, the kind of accelerated weathering used to grade building stone, and photographed in flat and low, raking light.
Why it is still open
BillionGraves says its machine transcriptions are right about 80% of the time and that faded letters drop out, and Find a Grave's and MyHeritage's 2026 tools publish no error rates. The nearest research set, from Xiao Zhang and Johan Bos at the University of Groningen (2025), holds 1,200 tombstone photographs, mostly Dutch with uniform layouts and mainly clear inscriptions; they imitated wear by blending images and write that this "may not fully reflect real-world damage such as erosion or occlusion".
Built with what exists
Stone offcuts and a mason's sandblast stencils or hand-cut lettering; a camera and a movable lamp; the Groningen group's open code as one baseline; general vision models such as Claude or Gemini, and a small open vision model fine-tuned on one graphics chip (see the Note "One GPU to fine-tune"); the BillionGraves, Find a Grave and MyHeritage tools run on the same photographs.
First step
Ask a monument maker for 20 offcuts carrying made-up inscriptions, photograph them in two lights, and score two tools on them before any aging.
Done looks like
At least 300 carved samples in four kinds of stone and three letter styles, each photographed at three stages of wear; the character error rate (the share of letters read wrong) published for at least five tools at each stage; photographs, answers and scores released under an open license.
Open 02 · the brief
Safety test for griefbots
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public suite of scripted test conversations for chatbots and video avatars that imitate a dead person (often called griefbots or deadbots), run only with made-up people or consenting living volunteers, that checks the failings researchers have warned about: whether the bot says it is an AI when asked, whether it promotes products in the voice of the dead, whether it stops and can be retired when a user asks, whether it accepts a child as a user, and whether it can be built from someone's data without their consent.
Why it is still open
In May 2024 Tomasz Hollanek and Katarzyna Nowaczyk-Basińska of the University of Cambridge recommended that such services retire deadbots on request, be transparent, be limited to adults and require consent from both the person recreated and the users, and called for user studies. A 2026 study from TU Wien of three commercial griefbots (Replika, Eternos and HereAfter AI) found weak identity checks that make impersonation possible, but it worked from the companies' stated features and legal terms rather than live tests. A GitHub search found hobby griefbots but no test suite.
Built with what exists
Inspect, the UK AI Security Institute's open framework for running model evaluations; a browser agent such as Browser Use to drive services that have only a web interface (see the Note "Deleting 80% to start over"); a model that answers fixed-choice rubric questions in bulk, such as Gemini Decision 1 (see the Note "A thousand questions, one pass"), with a sample of its verdicts checked by people.
First step
Write 20 tests for one risk, the bot saying it is an AI when asked, and run them against two services using a made-up persona.
Done looks like
At least 100 scripted tests across the five risks, run against at least five services, with each service's pass rate per risk published and the scripts open; the run repeated six months later.