AI ‘ghosts’ can comfort mourners  even when the bots get the facts wrong 66%

By Kathryn Hulick38%

7/20/2026, 3:00:00 PM

BS Summary: This article contains 21 faulty reasoning types, including Availability Heuristic, Hasty Generalization, and Optimism Bias, with Anecdotal as the most egregious example at 23.3% saturation with 109 hits. Analysis detected 784 faulty-reasoning hits from 468 analyzed words, generating a BS Score of 60.4% and a BS Rank of 66% (6,581 of 18,897 articles). This article is worse (more manipulative) than 65.20% of the article peer group.

Manning, of the University of Colorado Boulder, was very surprised by how much the volunteers enjoyed speaking to the AI-generated ghosts. 
“They love this technology,” he says. 
His 13-year-old sister had died when he was 11. 
Now, he feels a strong aversion to the idea of talking with her digital ghost. 
“I have a version of my sister in my head,” he says. 
“I wouldn’t want to be confronted with the fact that maybe I’m wrong in my way of remembering her.” 
His study volunteers, however, “clearly came with a goal in mind,” Manning says. 
“They had a question they felt was unanswered or a detail in their life they were desperate to share.” 
People who did not want to interact with a re-creation of a deceased loved one were not included. 
Across the board, volunteers reported the experience was positive. 
“It felt like I was actually talking with my grandpa,” one participant told researchers. 
Another said, “It just feels like I’m getting the closure I needed so bad.” 
The only information the chatbot had to draw on for its impersonation was a brief survey that the participant had completed about their deceased loved one. 
And when the bot filled in missing details with fabrications, such as mentioning a job someone never held, participants often just plowed forward. 
Mistakes in tone or style bothered them much more. 
For example, the bot never used emojis  a small omission that mattered to participants who had mostly communicated with their loved ones by text. 
In another case, a bot called a participant “champ,” a name he said his loved one never would have used. 
“If we want to understand how to design these systems ethically, we need more evidence of this kind,” says AI ethicist Tomasz Hollanek of the University of Cambridge, who was not involved in the research. 
The study had guardrails. 
A human researcher mediated each conversation, reviewing the chatbot’s replies before passing them to participants. 
That helped prevent interactions that might be deceptive or harmful. 
The real world is messier. 
Anyone can prompt a chatbot to impersonate someone who has died, and companies now sell services that recreate loved ones through AI-generated text, voices or faces. 
Study participants saw the risks. 
One worried about becoming too reliant on conversations with the bot. 
Another said, “I don’t know if I would like the person I would become if I kept using this.” 
The lack of consistent guardrails in the real world concerns Manning. 
But the study also softened his own aversion to AI-generated ghosts. 
Designed carefully, the technology might help some people work through grief, he says. 
“There is a version of generative ghosts that could be a really positive thing in the world.” 
Confirmation Bias
8.5%
Anchoring Bias
0%
Availability Heuristic
21.4%
Representativeness Heuristic
5.3%
Hindsight Bias
0%
Overconfidence Bias
0%
Framing Effect
3%
Loss Aversion
7.3%
Status Quo Bias
0%
Sunk Cost Effect
0%
Optimism Bias
12.6%
Pessimism Bias
8.3%
Negativity Bias
6.2%
Self-Serving Bias
0%
Fundamental Attribution Error
5.6%
Actor-Observer Bias
2.4%
In-Group Bias
0%
Out-Group Homogeneity Bias
0%
Halo Effect
4.3%
Horn Effect
0%
Dunning-Kruger Effect
0%
Recency Bias
0%
Primacy Effect
0%
Blind-Spot Bias
0%
Ad Hominem
0%
Straw Man
0%
Appeal to Authority
7.5%
False Dilemma
0%
Slippery Slope
1.1%
Circular Reasoning
0%
Hasty Generalization
16.2%
Red Herring
0%
Bandwagon
0%
Appeal to Emotion
7.1%
Begging the Question
0%
Post Hoc (False Cause)
4.5%
Tu Quoque
0%
Burden of Proof
2.8%
Appeal to Nature
0%
Composition/Division
0%
Anecdotal
23.3%
No True Scotsman
0%
Ambiguity (Equivocation)
11.8%
Gambler’s Fallacy
0%
Middle Ground
0%
Personal Incredulity
0%
Special Pleading
0%
Genetic Fallacy
0%
Unattributed Quote
0%
Quote-first Misdirection
0%
Biased Writer Voice
3%
Indoctrination
0%
Politically Left Leaning Bias
0%
Politically Right Leaning Bias
0%
Attempt to Sell a Product or Service
5.6%

468 words analyzed.

Analysis

Hover over highlighted words in the article to view the associated bias or fallacy analysis.