A Tokyo‑based startup, EmoMetric Labs, announced on Friday the joint development of a new algorithm capable of detecting “regret signatures” in typing speed. The project, conducted in collaboration with the Tokyo Institute for Human‑Computer Sentiment, aims to quantify emotional hesitation in digital communication with what the company describes as “uncomfortable precision.” EmoMetric Labs, founded three years ago and widely regarded as one of Japan’s few potential unicorns, has built its reputation on the belief that emotional states can be measured as reliably as temperature or humidity, an idea that has attracted both significant investment and quiet concern.
The algorithm, named HesitaType, analyzes micro‑variations in keystroke timing to determine whether a user is experiencing remorse, second thoughts, or the specific emotional turbulence associated with writing to a former romantic partner. While the system was originally intended for corporate use, flagging emails that employees might later wish they hadn’t sent, early trials revealed a striking anomaly: HesitaType achieved its highest accuracy when analyzing emails addressed to exes.
“We didn’t design it to do that. It simply gravitated toward the strongest available signal,” said Dr. Nagisa Kishibe, founder and CEO of EmoMetric Labs. Kishibe, who holds a PhD in computational affective science, spoke with the calm confidence of someone who has spent years quantifying emotions he no longer personally experiences. “Typing speed becomes unusually unstable when people attempt to sound casual while suppressing unresolved feelings. The model recognized that pattern immediately.”
Kishibe emphasized that the algorithm’s performance in ex‑related communication was “not intentional,” adding, “The model is unbiased. It just understands regret better than most people do.” He declined to comment on whether his own typing data was included in the training set, noting only that he has been “fully focused on work” since a breakup two years ago. Investors familiar with the company’s early days have suggested, off the record, that the breakup may have accelerated the startup’s pivot toward emotional analytics, though Kishibe has consistently framed the shift as “market‑driven.”
According to the research team, HesitaType identifies regret through a combination of hesitation patterns, micro‑pauses, and subtle fluctuations in keystroke rhythm. One common indicator is the extended pause, often exceeding 1.8 seconds, after typing the word “Hey,” a moment researchers describe as “the emotional cliff.” Another is the rapid backspacing that follows phrases like “Just wanted to check in,” which the algorithm interprets as a sign of internal conflict. The system also detects a measurable drop in what the team calls “confidence velocity” when users debate whether to include an emoji, particularly the neutral smile. In several cases, HesitaType identified regret before the user had typed anything beyond the recipient’s name. “The emotional signature was already present,” Kishibe explained. “The model detected anticipatory remorse, which is common in certain interpersonal histories.”
Cultural differences also emerged during testing. Japanese participants displayed a unique pattern of “polite regret,” characterized by extremely consistent typing speed punctuated by sudden, catastrophic bursts of backspacing. “It’s a very Japanese form of emotional compression,” Kishibe said. “The remorse is neatly packaged, but the algorithm can still see the seams.” Researchers noted that this pattern was especially pronounced in emails containing honorifics, where users attempted to maintain formality while navigating unresolved emotional terrain.
Tech companies have already expressed interest in licensing HesitaType. One major email provider is testing a feature that warns users when they are about to send a message they will later describe as “a mistake,” “a lapse in judgment,” or “something I shouldn’t have sent at 1:14 a.m.” The prototype displays a pop‑up reading: “This email contains detectable regret. Proceed anyway?” Early testers reported that the warning appeared with unsettling frequency, sometimes before they had finished typing the subject line.
User reactions have been mixed. Some appreciated the intervention, while others found it intrusive. One tester reported that the system triggered repeatedly when writing to his supervisor, despite the messages being strictly work‑related. “That’s not a bug,” Kishibe clarified. “That’s just your relationship with your supervisor.” Another participant complained that the algorithm flagged every message she attempted to send after midnight, regardless of content. The research team later confirmed that late‑night typing correlates strongly with “ambient regret,” a baseline emotional state the model is still learning to distinguish from ordinary fatigue.
EmoMetric Labs is now exploring whether HesitaType can be adapted to detect other emotional states, such as denial, misplaced confidence, or the specific panic associated with accidentally hitting reply‑all. However, Kishibe acknowledged that the algorithm may never fully overcome its tendency toward ex‑related communication. “It’s simply where the strongest signals are,” he said. “Human regret is complex, but emails to exes are its most concentrated form.”
