Letter review
Template and AI-drafted recommendation letters: can USCIS tell?
The Department of Homeland Security inventory of AI use cases lists 28 USCIS entries, and none is described as detecting text written by artificial intelligence. On December 30, 2013, the Administrative Appeals Office (AAO) gave little weight to letters supporting a national interest waiver petition because their language was "similar, and at times, identical" and did not appear to be the authors' own.

Can USCIS tell if a recommendation letter was written by AI?
The DHS inventory of USCIS AI use cases lists 28 entries, and its change log is dated January 15, 2025. None is described as detecting AI-generated text. One deployed model compares sentences across documents submitted to USCIS.
The Department of Homeland Security (DHS) lists the uses of artificial intelligence at USCIS on its AI use case inventory page for USCIS. The page has 28 entries: 6 in pre-deployment, 12 deployed and 10 inactive. Its change log has one entry, dated January 15, 2025. No entry is described as detecting text generated by artificial intelligence.
One deployed entry compares texts: DHS-130, the Text Analytics Data Science Sentence Similarity Model. In the words of the inventory, it "identifies similarity and relevancy in text to determine potential matches in documents submitted into the system." The model scans the narrative sections of applications for common language patterns, and trained staff decide whether a pattern points to possible fraud or to a national security or public safety concern. The entry states that the model "does not make any determinations or decisions" and marks it as rights-impacting.
A second text tool, DHS-13 Asylum Text Analytics, used machine learning and data graphing to "identify plagiarism-based fraud in applications for asylum status and for the withholding of removal." The inventory lists it as inactive. The deployed ELIS Evidence Classifier Machine Learning (ML) Tagging Solution (DHS-16) labels each scanned document with an evidence type for the adjudicator or leaves it unlabeled when it does not recognize the type. Under DHS-366, Large Language Models for an Officer Training Tool, generative AI produces training material for Refugee, Asylum, and International Operations officers.
Can a trained reader tell an AI letter from a human one?
In a study published in December 2023, 32 academic physicians with experience reviewing recommendation letters had a mean accuracy of 59.4% in telling human-written letters from AI-written ones. The authors found they could not do it reliably.
The study, "Brain versus bot: Distinguishing letters of recommendation authored by humans compared with artificial intelligence," appeared in AEM Education and Training, volume 7, issue 6 (DOI 10.1002/aet2.10924). Carl Preiksaitis and five co-authors used a single-blind quasi-experimental design. The letters supported promotion to associate professor in academic medicine, and the readers were academic physicians with experience reviewing such letters.
Mean accuracy was 59.4%. Neither the readers' certainty nor the time they spent on a letter had a significant effect on accuracy. Letters that readers took for human-written received higher ratings for quality and persuasiveness. The conclusion of the authors reads: "Participants were unable to reliably differentiate between human- and AI-authored LORs for promotion."
No immigration petition was part of the study.
What did the AAO do with letters that used the same wording?
On December 30, 2013, the AAO found letters supporting an NIW petition "similar, and at times, identical," wrote that the similarities suggest the language is not the authors' own, and held that the letters possess little credibility or probative value.
The petitioner, an early childhood multicultural education administrator, filed an immigrant petition under section 203(b)(2) of the Immigration and Nationality Act as a person of exceptional ability and asked for a national interest waiver (NIW) of the job offer requirement. The case came from the Texas Service Center. After an AAO decision of July 18, 2013, she filed a motion to reconsider, and in the decision on that motion, dated December 30, 2013, the AAO returned to the reference letters.
The AAO wrote that their language was "similar, and at times, identical," and that "these similarities across the various letters suggest that the language in the letters is not the authors' own." It relied on two asylum cases from the Second Circuit. In Surinder Singh v. Board of Immigration Appeals, 438 F.3d 145 (2d Cir. 2006), the court upheld an immigration judge's adverse credibility finding based in part on the similarity of some of the affidavits. In Mei Chai Ye v. U.S. Dept. of Justice, 489 F.3d 517 (2d Cir. 2007), the court concluded that a judge may reasonably infer that the applicant is the common source of "strikingly similar affidavits."
Counsel answered that the petitioner had given her supporters "draft sample letters" as a courtesy and that "each writer was free to edit, add to, or delete any part of the sample letter." In counsel's view, the repeated language emphasized the petitioner's abilities. The AAO held that "because the letters appear to have been drafted by someone other than the purported authors, the letters possess little credibility or probative value."
Citing Matter of Chawathe, 25 I&N Dec. 369, 376 (AAO 2010), the AAO wrote that the truth "is to be determined not by the quantity of evidence alone but by its quality." It also cited Matter of Caron International, 19 I&N Dec. 791, 795 (Comm'r 1988): USCIS may use statements submitted as expert testimony as advisory opinions and is "ultimately responsible for making the final determination regarding an alien's eligibility." The motion was dismissed, and the petition remained denied.
Can the applicant or the attorney give the author a draft?
The USCIS Policy Manual passage on letters does not say who drafts the text. The Yale Office of International Students and Scholars tells O-1 applicants that "the letters should be written by the recommenders."
Volume 6, Part F, Chapter 2 of the USCIS Policy Manual has a passage headed "Letters of Support." The letters "should explain in specific terms" why the witnesses hold their view of the applicant, and the statements in them "should be corroborated by documentary evidence in the record." The passage says nothing about who prepares the text. In the 2013 case the AAO gave little weight to letters that counsel said had grown out of the petitioner's "draft sample letters," although each writer was, in counsel's words, free to edit them.
The Yale Office of International Students and Scholars (Yale OISS) writes on its page O-1 Letters of Recommendation: "The letters should be written by the recommenders." The page states that "Immigration has denied applications based on their sense that the letters were not written by the recommenders" and that USCIS looks for evidence that the expert recommender did not write the letter, "e.g. identical phrases, sequence and/or format." It adds: "They have been known in these cases to summarily dismiss all letters as not credible." Yale OISS itself "is not in a position to draft or edit letters" and only gives feedback on them.
In that list "sequence" and "format" stand next to "identical phrases" with "and/or" between them, so a set of letters with no sentence in common still fits it when every letter follows one order of paragraphs and one layout.
What must a letter contain that a template cannot supply?
The author's own basis of knowledge, the specific reasons for the author's view and the author's relationship with the applicant. The Policy Manual names all three, and they differ between authors.
For original contributions of major significance, the same chapter says that submitted letters "should specifically describe the person's contribution and its significance to the field and should also set forth the basis of the writer's knowledge and expertise." Each author has a different basis: the applicant's papers the author has read, or the project where the author used the applicant's method. The officer should also consider the relationship or affiliation between the applicant and the witness "when evaluating the significance of witnesses' statements," and that relationship is different for every author. A template written before the author agreed to sign contains neither the basis nor the relationship.
The chapter also states that letters "that merely reiterate USCIS' definitions relating to this classification" are "generally not persuasive." Letters that make only general assertions about the applicant at most indicate "a competent, respected figure" in the field. A template built on the wording of the regulation puts those definitions into every letter of the set.
What happens when an AI draft states something that is not true?
In Mata v. Avianca a federal judge imposed a $5,000 penalty for filing case citations that ChatGPT had invented. Under INA 212(a)(6)(C)(i), a person who seeks a visa or other benefit by willfully misrepresenting a material fact is inadmissible.
On June 22, 2023, Judge P. Kevin Castel of the U.S. District Court for the Southern District of New York issued an opinion and order on sanctions in Mata v. Avianca, Inc., No. 22-cv-1461 (PKC). The respondents, lawyers Peter LoDuca and Steven A. Schwartz and the firm Levidow, Levidow & Oberman P.C., had "submitted non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT." They acknowledged that six of the cited decisions were fabricated. Under Rule 11 the court imposed a $5,000 penalty, jointly and severally, payable within 14 days, and ordered letters to the client and to each judge falsely named as the author of a fake opinion. It found bad faith on the part of the two lawyers.
That case concerned a court filing, and the false items were case citations. A recommendation letter has its own items that can be checked: citation counts, grant numbers, awards, dates of employment. Under the Policy Manual the statements in letters "should be corroborated by documentary evidence in the record," and the officer can compare each such item with the exhibits.
The consequence of a false statement is set by statute. INA 212(a)(6)(C)(i), quoted in 9 FAM 302.9-4(A) of the State Department's Foreign Affairs Manual, covers "any alien who by fraud or willfully misrepresenting a material fact seeks to procure (or sought to procure or has procured)" a visa, other documentation, admission or "some other benefit provided under" the Act. Under the manual, a misrepresentation is material if the applicant is ineligible on the true facts or if it "tends to shut off a line of inquiry which is relevant to the alien's eligibility." The same manual says that a misrepresentation can be made in an oral interview, in written applications "or by submitting evidence containing false information." An applicant who goes through an attorney or another third party remains liable for misrepresentations made by that agent if the applicant was aware of the action being taken in furtherance of the application. Saying that someone else advised the action is no defense, unless the applicant lacked the capacity to exercise judgment.
How do you check a set of letters for template signs before filing?
Put the letters side by side. Mark repeated sentences, the same order of paragraphs and the same layout, then match every figure, date and award in each letter to an exhibit in the record.
The check uses the signs named by the AAO in 2013 and by Yale OISS, together with the corroboration rule of the Policy Manual. It is done on the final versions that the authors will sign.
- Put all letters of the petition next to each other and mark every sentence that appears in two or more of them.
- Compare the order of paragraphs and the layout of each letter with the others, because Yale OISS names sequence and format along with identical phrases.
- Find each sentence that restates the regulation or the Policy Manual and ask the author which fact stands behind it.
- Check that each author states how he or she knows the applicant's work and what the relationship is.
- Match every figure, date, grant, award and publication named in a letter to an exhibit, and remove what has no exhibit.
- Send each author the final text before signature and keep the author's written confirmation that the text states the author's own opinion.
What does an IAE letter review check?
A reviewer from the IAE registry, independent of the letter's provider, checks the letter page by page against the criteria of the visa category and returns a signed 3-5 page review with one verdict: file, revise or redo.
The IAE expert letter review is an independent check of an expert opinion letter or a credential evaluation before filing with USCIS. The service page names template text, declarative conclusions and an author without relevant credentials among the weak points the reviewer looks for. Every section of the letter is checked against the criteria of the visa category, and the credentials and relevance of the signing expert are assessed separately.
The review lists specific weak points with a way to strengthen each, and the applicant and the attorney revise the letter from that list. It ends with one of three verdicts, each with reasoning: file, revise or redo. If the letter turns out to be strong, the verdict is to file, and the letter goes into the package without changes. The review runs 3-5 pages and is signed by an IAE registry member in the letter's field.
IAE reviews letters from any company. The reviewer is independent of the provider of the document under review and confirms this in writing, and the materials are kept confidential. After an RFE, the same review answers whether the officer's findings stem from a weak expert letter. It can also compare two letters, or samples of work, from different companies.
To order, send the letter or the evaluation with a short description of the petition. The Secretariat confirms the fee, the turnaround and the reviewer within one business day, and the fee depends on the visa category and the volume of materials. The standard turnaround is 3 business days from payment and receipt of all materials, or 1 business day on the rush option. IAE is not a law firm, and USCIS decides the petition.
Sources
- DHS, AI Use Case Inventory: U.S. Citizenship and Immigration Services (change log entry of January 15, 2025)
- Preiksaitis C. et al., Brain versus bot: Distinguishing letters of recommendation authored by humans compared with artificial intelligence, AEM Education and Training 7(6), 2023 (Duke Scholars record)
- Brain versus bot, abstract (Toronto Metropolitan University repository)
- AAO decision on motion to reconsider, immigrant petition under INA 203(b)(2), exceptional ability and national interest waiver (Dec. 30, 2013)
- USCIS Policy Manual, Volume 6, Part F, Chapter 2, Extraordinary Ability
- Yale OISS, O-1 Letters of Recommendation
- Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), Opinion and Order on Sanctions (S.D.N.Y. June 22, 2023)
- U.S. Department of State, Foreign Affairs Manual, 9 FAM 302.9 (INA 212(a)(6)(C)(i), fraud and misrepresentation)
Read also
All articles
Indian three-year bachelor's plus two-year master's: does USCIS count it as an advanced degree for EB-2?
In AAO decisions of 2017 and 2019 an Indian master's degree was placed at the level of a U.S. bachelor's degree. In 2017 the AAO cited EDGE. In 2019
Foreign medical degree (MBBS or MD) for USCIS: when is it an advanced degree, and what must the evaluation show?
A U.S. medical degree is an advanced degree for USCIS because U.S. medical schools admit students after a bachelor's. Many foreign schools admit them
USCIS cited AACRAO EDGE against your credential evaluation: what happens next?
When a credential evaluation and the AACRAO EDGE database disagree, USCIS may follow EDGE. In Viraj v. U.S. Attorney General (11th Cir. 2014)Expert letter review
Check the letter before the officer does
The standard turnaround is 3 business days, the rush option 1 business day. The fee depends on the visa category and the volume of materials; we confirm the fee, the turnaround and the reviewer within one business day.

