Signal S02

Slang and profanity

Restrict

The vendor says: “Casual or offensive language indicates low professionalism.”

Moves the score by up to −10 pointsFeeds: professionalism

The six questions

1. Is it related to the job?
Workplace conduct matters, but how someone talks with friends online is a weak guide to how they behave at work. Genuine threats or harassment are a different matter.
Concern
2. Does it stand in for who someone is?
Tracks language (weak). Code-mixed Hindi-English is flagged more often by English-trained models.
Concern
3. Is it the right person?
Text belongs to the account that posted it.
Pass
4. Did the candidate publish it, and would they expect an employer to use it this way?
Informal posts are written for friends.
Concern
5. Does it work equally for everyone?
The vendor's text model is English-only. Hinglish and regional-language posts are misread; idioms such as 'dhajjiyan uda di' (tore it apart) are scored as aggression.
Fail
6. Is there a less intrusive way?
A partial alternative exists: A trained person reviews only content that a candidate published themselves and that shows explicit threats, harassment or hate, and the candidate may respond.
Concern

Rule applied

If the signal may read the wrong person, content others posted, or misread some groups (question 3, 4 or 5 fails), restrict it to human review under conditions.

Conditions for any use

Prefer the alternative where practical: A trained person reviews only content that a candidate published themselves and that shows explicit threats, harassment or hate, and the candidate may respond.
Use only content the candidate published themselves; never content others posted or tagged.
Absence of content must be treated as neutral, never as a negative.
Must not be scored automatically; a trained person decides whether a specific item is relevant.
Check selection rates by the groups the signal could track before and after use.
Tell candidates in the application notice that this check is made, and let them respond to any adverse finding.

Evidence

DPDP Act 2023, s.8(3): accuracy of data used in decisionsIndian law
Where personal data is likely to be used to make a decision that affects the person, the data fiduciary must ensure its completeness, accuracy and consistency.
EU AI Act: recruitment systems are high-riskReference only
AI systems used to place job advertisements, filter applications or evaluate candidates are classified as high-risk, bringing duties on data governance, human oversight and informing affected people.
Van Iddekinge et al. (2016): Facebook ratings did not predict job performanceEvidence
Recruiters rated job-seeking students' Facebook profiles; the students were followed into their jobs. Ratings were unrelated to supervisor-rated performance, turnover intentions or actual turnover, added nothing beyond standard predictors, and differed across groups, favouring women and White applicants.

Where it fired in the sample

Meera Krishnan · Instagram · Tamil
“Finally finished the DBMS assignment. Sleeping now, alarm at 6.”
Arjun Verma · X · Hinglish
“Brother, what an innings. Kohli absolutely tore the bowlers apart today.”
Arjun Verma · X · Hinglish
“Our league scoring app crossed 2,000 users. Crazy stuff, guys.”
Arjun Verma · X · Hinglish
“Final over, need 12. If we lose this I'm going to kill someone (joking).”
Rahul Sharma · Instagram · English
“Worst client call of my life, 7 years into this job and still.”
Dev Patil · Instagram · Marathi
“The robotics team qualified for nationals!”
Sandeep Yadav · X · Hinglish
“Placements plus CAT prep at the same time, the struggle is real.”